mirror of
https://github.com/tinygrad/tinygrad.git
synced 2026-01-07 22:23:55 -05:00
3019 lines
266 KiB
ArmAsm
3019 lines
266 KiB
ArmAsm
prologue: // 0000000000001600
|
|
s_load_b64 s[28:29], s[0:1], null // 000000001600: F4040700 F8000000
|
|
s_load_b64 s[34:35], s[0:1], 0x8 // 000000001608: F4040880 F8000008
|
|
s_load_b64 s[32:33], s[0:1], 0x10 // 000000001610: F4040800 F8000010
|
|
s_waitcnt lgkmcnt(0) // 000000001618: BF89FC07
|
|
s_mov_b32 s47, 1 // 00000000161C: BEAF0081
|
|
s_mov_b32 s48, 0 // 000000001620: BEB00080
|
|
s_mov_b32 s49, 0x2200001 // 000000001624: BEB100FF 02200001
|
|
s_mov_b32 s11, 0xc010008 // 00000000162C: BE8B00FF 0C010008
|
|
s_mov_b32 s50, 0x739 // 000000001634: BEB200FF 00000739
|
|
s_mov_b32 s24, 4096 // 00000000163C: BE9800FF 00001000
|
|
s_mov_b32 s25, s24 // 000000001644: BE990018
|
|
s_mov_b32 s26, 1 // 000000001648: BE9A0081
|
|
s_mov_b32 s27, s24 // 00000000164C: BE9B0018
|
|
s_mov_b32 s36, s24 // 000000001650: BEA40018
|
|
s_mov_b32 s37, 0 // 000000001654: BEA50080
|
|
s_mov_b32 s38, s24 // 000000001658: BEA60018
|
|
s_mov_b32 s39, 0 // 00000000165C: BEA70080
|
|
s_mov_b32 s40, s24 // 000000001660: BEA80018
|
|
s_mov_b32 s41, 0 // 000000001664: BEA90080
|
|
s_mov_b32 s42, s24 // 000000001668: BEAA0018
|
|
s_mov_b32 s43, 0 // 00000000166C: BEAB0080
|
|
s_and_b32 s10, s49, 0xffff0000 // 000000001670: 8B0AFF31 FFFF0000
|
|
s_lshr_b32 s10, s10, 16 // 000000001678: 850A900A
|
|
s_and_b32 s46, s49, 0xffff // 00000000167C: 8B2EFF31 0000FFFF
|
|
s_mov_b32 s5, s48 // 000000001684: BE850030
|
|
s_mov_b32 m0, 0x7680 // 000000001688: BEFD00FF 00007680
|
|
v_mov_b32_e32 v254, v0 // 000000001690: 7FFC0300
|
|
s_mov_b32 vcc_hi, 0 // 000000001694: BEEB0080
|
|
s_lshr_b32 s56, s11, 16 // 000000001698: 8538900B
|
|
s_ctz_i32_b32 s56, s56 // 00000000169C: BEB80838
|
|
s_lshr_b32 s57, s11, 22 // 0000000016A0: 8539960B
|
|
s_cmp_gt_i32 s56, 0 // 0000000016A4: BF028038
|
|
v_and_b32_e32 v1, 31, v254 // 0000000016A8: 3603FC9F
|
|
v_and_b32_e32 v0, 15, v1 // 0000000016AC: 3600028F
|
|
v_lshrrev_b32_e32 v4, 5, v254 // 0000000016B0: 3209FC85
|
|
v_and_b32_e32 v4, 1, v4 // 0000000016B4: 36080881
|
|
v_lshl_add_u32 v0, v4, 4, v0 // 0000000016B8: D6460000 04010904
|
|
v_and_b32_e32 v2, 31, v254 // 0000000016C0: 3605FC9F
|
|
v_and_b32_e32 v1, 15, v2 // 0000000016C4: 3602048F
|
|
v_lshlrev_b32_e32 v1, 5, v1 // 0000000016C8: 30020285
|
|
v_lshrrev_b32_e32 v3, 6, v254 // 0000000016CC: 3207FC86
|
|
v_and_b32_e32 v3, 1, v3 // 0000000016D0: 36060681
|
|
v_lshl_add_u32 v1, v3, 9, v1 // 0000000016D4: D6460001 04051303
|
|
v_lshrrev_b32_e32 v2, 5, v254 // 0000000016DC: 3205FC85
|
|
v_lshrrev_b32_e32 v2, 2, v2 // 0000000016E0: 32040482
|
|
s_mov_b32 s49, 0xc00 // 0000000016E4: BEB100FF 00000C00
|
|
v_mul_lo_u32 v2, s49, v2 // 0000000016EC: D72C0002 00020431
|
|
v_add_lshl_u32 v80, v2, v0, 1 // 0000000016F4: D6470050 02060102
|
|
v_mov_b32_e32 v4, 0x2aaaab // 0000000016FC: 7E0802FF 002AAAAB
|
|
v_mul_hi_u32 v5, v80, v4 // 000000001704: D72D0005 00020950
|
|
v_mul_lo_u32 v4, v80, v4 // 00000000170C: D72C0004 00020950
|
|
v_lshrrev_b64 v[4:5], 33, v[4:5] // 000000001714: D73D0004 000208A1
|
|
v_mov_b32_e32 v3, v4 // 00000000171C: 7E060304
|
|
v_lshl_add_u32 v80, v3, 5, v80 // 000000001720: D6460050 05410B03
|
|
v_lshrrev_b32_e32 v0, 5, v254 // 000000001728: 3201FC85
|
|
v_lshrrev_b32_e32 v0, 2, v0 // 00000000172C: 32000082
|
|
s_mov_b32 s49, 32 // 000000001730: BEB100A0
|
|
v_mul_lo_u32 v0, s49, v0 // 000000001734: D72C0000 00020031
|
|
v_add_lshl_u32 v81, v0, v1, 1 // 00000000173C: D6470051 02060300
|
|
v_lshrrev_b32_e32 v2, 7, v81 // 000000001744: 3204A287
|
|
v_lshl_add_u32 v81, v2, 5, v81 // 000000001748: D6460051 05450B02
|
|
v_add_co_u32 v81, vcc_lo, 0x1880, v81 // 000000001750: D7006A51 0002A2FF 00001880
|
|
v_lshrrev_b32_e32 v1, 2, v254 // 00000000175C: 3203FC82
|
|
v_and_b32_e32 v0, 3, v254 // 000000001760: 3601FC83
|
|
v_lshlrev_b32_e32 v0, 3, v0 // 000000001764: 30000083
|
|
v_mov_b32_e32 v4, v1 // 000000001768: 7E080301
|
|
v_lshrrev_b32_e32 v2, 2, v254 // 00000000176C: 3205FC82
|
|
v_and_b32_e32 v3, 3, v254 // 000000001770: 3607FC83
|
|
v_lshlrev_b32_e32 v3, 3, v3 // 000000001774: 30060683
|
|
v_mov_b32_e32 v5, v3 // 000000001778: 7E0A0303
|
|
v_mul_u32_u24_e32 v78, 0x60, v4 // 00000000177C: 169C08FF 00000060
|
|
v_add_lshl_u32 v78, v0, v78, 1 // 000000001784: D647004E 02069D00
|
|
v_mov_b32_e32 v6, 0x2aaaab // 00000000178C: 7E0C02FF 002AAAAB
|
|
v_mul_hi_u32 v7, v78, v6 // 000000001794: D72D0007 00020D4E
|
|
v_mul_lo_u32 v6, v78, v6 // 00000000179C: D72C0006 00020D4E
|
|
v_lshrrev_b64 v[6:7], 33, v[6:7] // 0000000017A4: D73D0006 00020CA1
|
|
v_mov_b32_e32 v6, v6 // 0000000017AC: 7E0C0306
|
|
v_lshl_add_u32 v78, v6, 5, v78 // 0000000017B0: D646004E 05390B06
|
|
v_mul_u32_u24_e32 v79, 32, v2 // 0000000017B8: 169E04A0
|
|
v_add_lshl_u32 v79, v5, v79, 1 // 0000000017BC: D647004F 02069F05
|
|
v_lshrrev_b32_e32 v6, 7, v79 // 0000000017C4: 320C9E87
|
|
v_lshl_add_u32 v79, v6, 5, v79 // 0000000017C8: D646004F 053D0B06
|
|
v_add_co_u32 v79, vcc_lo, 0x1880, v79 // 0000000017D0: D7006A4F 00029EFF 00001880
|
|
s_waitcnt lgkmcnt(0) // 0000000017DC: BF89FC07
|
|
v_mov_b32_e32 v8, 0x60 // 0000000017E0: 7E1002FF 00000060
|
|
v_mov_b32_e32 v7, s24 // 0000000017E8: 7E0E0218
|
|
v_cvt_f32_u32_e32 v6, v8 // 0000000017EC: 7E0C0D08
|
|
v_rcp_iflag_f32_e32 v6, v6 // 0000000017F0: 7E0C5706
|
|
v_cvt_f32_u32_e32 v9, v7 // 0000000017F4: 7E120D07
|
|
v_mul_f32_e32 v6, v6, v9 // 0000000017F8: 100C1306
|
|
v_cvt_u32_f32_e32 v6, v6 // 0000000017FC: 7E0C0F06
|
|
v_mul_u32_u24_e32 v9, v6, v8 // 000000001800: 16121106
|
|
v_sub_nc_u32_e32 v9, v7, v9 // 000000001804: 4C121307
|
|
v_cmp_ne_u32_e64 vcc_lo, v9, 0 // 000000001808: D44D006A 00010109
|
|
v_add_co_ci_u32_e64 v6, vcc_lo, v6, 0, vcc_lo // 000000001810: D5206A06 01A90106
|
|
v_mov_b32_e32 v8, 0x60 // 000000001818: 7E1002FF 00000060
|
|
v_mov_b32_e32 v7, s25 // 000000001820: 7E0E0219
|
|
v_readfirstlane_b32 s14, v6 // 000000001824: 7E1C0506
|
|
v_cvt_f32_u32_e32 v6, v8 // 000000001828: 7E0C0D08
|
|
v_rcp_iflag_f32_e32 v6, v6 // 00000000182C: 7E0C5706
|
|
v_cvt_f32_u32_e32 v9, v7 // 000000001830: 7E120D07
|
|
v_mul_f32_e32 v6, v6, v9 // 000000001834: 100C1306
|
|
v_cvt_u32_f32_e32 v6, v6 // 000000001838: 7E0C0F06
|
|
v_mul_u32_u24_e32 v9, v6, v8 // 00000000183C: 16121106
|
|
v_sub_nc_u32_e32 v9, v7, v9 // 000000001840: 4C121307
|
|
v_cmp_ne_u32_e64 vcc_lo, v9, 0 // 000000001844: D44D006A 00010109
|
|
v_add_co_ci_u32_e64 v6, vcc_lo, v6, 0, vcc_lo // 00000000184C: D5206A06 01A90106
|
|
v_readfirstlane_b32 s15, v6 // 000000001854: 7E1E0506
|
|
s_mul_i32 s48, s14, s15 // 000000001858: 96300F0E
|
|
s_and_b32 s49, s46, 0x3fff // 00000000185C: 8B31FF2E 00003FFF
|
|
s_mul_i32 s48, s48, s49 // 000000001864: 96303130
|
|
v_cvt_f32_u32_e32 v6, s48 // 000000001868: 7E0C0C30
|
|
v_rcp_iflag_f32_e32 v6, v6 // 00000000186C: 7E0C5706
|
|
v_cvt_f32_u32_e32 v7, s2 // 000000001870: 7E0E0C02
|
|
v_mul_f32_e32 v6, v6, v7 // 000000001874: 100C0F06
|
|
v_cvt_u32_f32_e32 v6, v6 // 000000001878: 7E0C0F06
|
|
v_mul_u32_u24_e64 v7, v6, s48 // 00000000187C: D50B0007 00006106
|
|
v_sub_nc_u32_e32 v7, s2, v7 // 000000001884: 4C0E0E02
|
|
v_cmp_eq_u32_e64 vcc_lo, v7, s48 // 000000001888: D44A006A 00006107
|
|
s_mov_b32 exec_lo, vcc_lo // 000000001890: BEFE006A
|
|
v_add_nc_u32_e32 v6, 1, v6 // 000000001894: 4A0C0C81
|
|
s_mov_b32 exec_lo, -1 // 000000001898: BEFE00C1
|
|
v_cmp_gt_u32_e64 vcc_lo, v7, s48 // 00000000189C: D44C006A 00006107
|
|
s_mov_b32 exec_lo, vcc_lo // 0000000018A4: BEFE006A
|
|
v_sub_nc_u32_e64 v6, v6, 1 // 0000000018A8: D5260006 00010306
|
|
s_mov_b32 exec_lo, -1 // 0000000018B0: BEFE00C1
|
|
v_readfirstlane_b32 s48, v6 // 0000000018B4: 7E600506
|
|
s_mov_b32 s4, s48 // 0000000018B8: BE840030
|
|
s_mul_i32 s48, s15, s14 // 0000000018BC: 96300E0F
|
|
s_mul_i32 s48, s48, s4 // 0000000018C0: 96300430
|
|
s_mul_i32 s48, s48, s49 // 0000000018C4: 96303130
|
|
s_sub_u32 s2, s2, s48 // 0000000018C8: 80823002
|
|
v_cvt_f32_u32_e32 v6, s14 // 0000000018CC: 7E0C0C0E
|
|
v_rcp_iflag_f32_e32 v6, v6 // 0000000018D0: 7E0C5706
|
|
v_cvt_f32_u32_e32 v7, s2 // 0000000018D4: 7E0E0C02
|
|
v_mul_f32_e32 v6, v6, v7 // 0000000018D8: 100C0F06
|
|
v_cvt_u32_f32_e32 v6, v6 // 0000000018DC: 7E0C0F06
|
|
v_mul_u32_u24_e64 v7, v6, s14 // 0000000018E0: D50B0007 00001D06
|
|
v_sub_nc_u32_e32 v7, s2, v7 // 0000000018E8: 4C0E0E02
|
|
v_cmp_eq_u32_e64 vcc_lo, v7, s14 // 0000000018EC: D44A006A 00001D07
|
|
s_mov_b32 exec_lo, vcc_lo // 0000000018F4: BEFE006A
|
|
v_add_nc_u32_e32 v6, 1, v6 // 0000000018F8: 4A0C0C81
|
|
s_mov_b32 exec_lo, -1 // 0000000018FC: BEFE00C1
|
|
v_cmp_gt_u32_e64 vcc_lo, v7, s14 // 000000001900: D44C006A 00001D07
|
|
s_mov_b32 exec_lo, vcc_lo // 000000001908: BEFE006A
|
|
v_sub_nc_u32_e64 v6, v6, 1 // 00000000190C: D5260006 00010306
|
|
s_mov_b32 exec_lo, -1 // 000000001914: BEFE00C1
|
|
v_readfirstlane_b32 s48, v6 // 000000001918: 7E600506
|
|
s_mov_b32 s3, s48 // 00000000191C: BE830030
|
|
s_mul_i32 s48, s3, s14 // 000000001920: 96300E03
|
|
s_sub_u32 s2, s2, s48 // 000000001924: 80823002
|
|
s_sub_u32 s32, s32, 16 // 000000001928: 80A09020
|
|
s_subb_u32 s33, s33, 0 // 00000000192C: 82A18021
|
|
s_sub_u32 s34, s34, 16 // 000000001930: 80A29022
|
|
s_subb_u32 s35, s35, 0 // 000000001934: 82A38023
|
|
s_mov_b64 s[6:7], 0 // 000000001938: BE860180
|
|
s_mov_b32 s8, 1 // 00000000193C: BE880081
|
|
s_mov_b32 s9, 1 // 000000001940: BE890081
|
|
s_sext_i32_i16 s11, s11 // 000000001944: BE8B0F0B
|
|
s_mov_b32 s11, s11 // 000000001948: BE8B000B
|
|
v_cvt_f32_u32_e32 v6, s11 // 00000000194C: 7E0C0C0B
|
|
v_rcp_iflag_f32_e32 v6, v6 // 000000001950: 7E0C5706
|
|
v_cvt_f32_u32_e32 v7, s3 // 000000001954: 7E0E0C03
|
|
v_mul_f32_e32 v6, v6, v7 // 000000001958: 100C0F06
|
|
v_cvt_u32_f32_e32 v6, v6 // 00000000195C: 7E0C0F06
|
|
v_mul_u32_u24_e64 v7, v6, s11 // 000000001960: D50B0007 00001706
|
|
v_sub_nc_u32_e32 v7, s3, v7 // 000000001968: 4C0E0E03
|
|
v_cmp_eq_u32_e64 vcc_lo, v7, s11 // 00000000196C: D44A006A 00001707
|
|
s_mov_b32 exec_lo, vcc_lo // 000000001974: BEFE006A
|
|
v_add_nc_u32_e32 v6, 1, v6 // 000000001978: 4A0C0C81
|
|
s_mov_b32 exec_lo, -1 // 00000000197C: BEFE00C1
|
|
v_cmp_gt_u32_e64 vcc_lo, v7, s11 // 000000001980: D44C006A 00001707
|
|
s_mov_b32 exec_lo, vcc_lo // 000000001988: BEFE006A
|
|
v_sub_nc_u32_e64 v6, v6, 1 // 00000000198C: D5260006 00010306
|
|
s_mov_b32 exec_lo, -1 // 000000001994: BEFE00C1
|
|
v_readfirstlane_b32 s68, v6 // 000000001998: 7E880506
|
|
s_mul_i32 s69, s68, s11 // 00000000199C: 96450B44
|
|
s_sub_u32 s69, s3, s69 // 0000000019A0: 80C54503
|
|
s_mul_i32 s69, s69, s14 // 0000000019A4: 96450E45
|
|
s_add_u32 s69, s69, s2 // 0000000019A8: 80450245
|
|
v_cvt_f32_u32_e32 v6, s11 // 0000000019AC: 7E0C0C0B
|
|
v_rcp_iflag_f32_e32 v6, v6 // 0000000019B0: 7E0C5706
|
|
v_cvt_f32_u32_e32 v7, s15 // 0000000019B4: 7E0E0C0F
|
|
v_mul_f32_e32 v6, v6, v7 // 0000000019B8: 100C0F06
|
|
v_cvt_u32_f32_e32 v6, v6 // 0000000019BC: 7E0C0F06
|
|
v_mul_u32_u24_e64 v7, v6, s11 // 0000000019C0: D50B0007 00001706
|
|
v_sub_nc_u32_e32 v7, s15, v7 // 0000000019C8: 4C0E0E0F
|
|
v_cmp_eq_u32_e64 vcc_lo, v7, s11 // 0000000019CC: D44A006A 00001707
|
|
s_mov_b32 exec_lo, vcc_lo // 0000000019D4: BEFE006A
|
|
v_add_nc_u32_e32 v6, 1, v6 // 0000000019D8: 4A0C0C81
|
|
s_mov_b32 exec_lo, -1 // 0000000019DC: BEFE00C1
|
|
v_cmp_gt_u32_e64 vcc_lo, v7, s11 // 0000000019E0: D44C006A 00001707
|
|
s_mov_b32 exec_lo, vcc_lo // 0000000019E8: BEFE006A
|
|
v_sub_nc_u32_e64 v6, v6, 1 // 0000000019EC: D5260006 00010306
|
|
s_mov_b32 exec_lo, -1 // 0000000019F4: BEFE00C1
|
|
v_readfirstlane_b32 s66, v6 // 0000000019F8: 7E840506
|
|
s_mul_i32 s67, s11, s66 // 0000000019FC: 9643420B
|
|
s_sub_u32 s67, s15, s67 // 000000001A00: 80C3430F
|
|
s_cmp_eq_u32 s67, 0 // 000000001A04: BF068043
|
|
s_cmov_b32 s67, s11 // 000000001A08: BEC3020B
|
|
s_cmp_ge_u32 s68, s66 // 000000001A0C: BF094244
|
|
s_cselect_b32 s66, s67, s11 // 000000001A10: 98420B43
|
|
v_cvt_f32_u32_e32 v6, s66 // 000000001A14: 7E0C0C42
|
|
v_rcp_iflag_f32_e32 v6, v6 // 000000001A18: 7E0C5706
|
|
v_cvt_f32_u32_e32 v7, s69 // 000000001A1C: 7E0E0C45
|
|
v_mul_f32_e32 v6, v6, v7 // 000000001A20: 100C0F06
|
|
v_cvt_u32_f32_e32 v6, v6 // 000000001A24: 7E0C0F06
|
|
v_mul_u32_u24_e64 v7, v6, s66 // 000000001A28: D50B0007 00008506
|
|
v_sub_nc_u32_e32 v7, s69, v7 // 000000001A30: 4C0E0E45
|
|
v_cmp_eq_u32_e64 vcc_lo, v7, s66 // 000000001A34: D44A006A 00008507
|
|
s_mov_b32 exec_lo, vcc_lo // 000000001A3C: BEFE006A
|
|
v_add_nc_u32_e32 v6, 1, v6 // 000000001A40: 4A0C0C81
|
|
v_mov_b32_e32 v7, 0 // 000000001A44: 7E0E0280
|
|
s_mov_b32 exec_lo, -1 // 000000001A48: BEFE00C1
|
|
v_cmp_gt_u32_e64 vcc_lo, v7, s66 // 000000001A4C: D44C006A 00008507
|
|
s_mov_b32 exec_lo, vcc_lo // 000000001A54: BEFE006A
|
|
v_sub_nc_u32_e64 v6, v6, 1 // 000000001A58: D5260006 00010306
|
|
v_mul_u32_u24_e64 v7, v6, s66 // 000000001A60: D50B0007 00008506
|
|
v_sub_nc_u32_e32 v7, s69, v7 // 000000001A68: 4C0E0E45
|
|
s_mov_b32 exec_lo, -1 // 000000001A6C: BEFE00C1
|
|
v_readfirstlane_b32 s2, v6 // 000000001A70: 7E040506
|
|
v_readfirstlane_b32 s3, v7 // 000000001A74: 7E060507
|
|
s_mul_i32 s3, s2, s66 // 000000001A78: 96034202
|
|
s_sub_u32 s3, s69, s3 // 000000001A7C: 80830345
|
|
s_mul_i32 s68, s68, s11 // 000000001A80: 96440B44
|
|
s_add_u32 s3, s3, s68 // 000000001A84: 80034403
|
|
v_mov_b32_e32 v6, v0 // 000000001A88: 7E0C0300
|
|
v_add_co_u32 v7, vcc_lo, 32, v6 // 000000001A8C: D7006A07 00020CA0
|
|
v_add_co_u32 v8, vcc_lo, 32, v7 // 000000001A94: D7006A08 00020EA0
|
|
v_mov_b32_e32 v9, v2 // 000000001A9C: 7E120302
|
|
v_add_co_u32 v10, vcc_lo, 32, v9 // 000000001AA0: D7006A0A 000212A0
|
|
v_add_co_u32 v11, vcc_lo, 32, v10 // 000000001AA8: D7006A0B 000214A0
|
|
v_mov_b32_e32 v12, v1 // 000000001AB0: 7E180301
|
|
v_mov_b32_e32 v13, v3 // 000000001AB4: 7E1A0303
|
|
s_mul_i32 s66, s2, 0x60 // 000000001AB8: 9642FF02 00000060
|
|
s_sub_u32 s66, s24, s66 // 000000001AC0: 80C24218
|
|
s_sub_u32 s66, s66, 8 // 000000001AC4: 80C28842
|
|
v_mov_b32_e32 v14, s66 // 000000001AC8: 7E1C0242
|
|
v_min_i32_e32 v6, v14, v6 // 000000001ACC: 220C0D0E
|
|
v_min_i32_e32 v7, v14, v7 // 000000001AD0: 220E0F0E
|
|
v_min_i32_e32 v8, v14, v8 // 000000001AD4: 2210110E
|
|
v_mul_lo_u32 v14, s40, v12 // 000000001AD8: D72C000E 00021828
|
|
v_add_co_u32 v72, vcc_lo, v6, v14 // 000000001AE0: D7006A48 00021D06
|
|
v_add_nc_u32_e32 v72, 8, v72 // 000000001AE8: 4A909088
|
|
v_lshlrev_b32_e32 v72, 1, v72 // 000000001AEC: 30909081
|
|
v_mul_lo_u32 v14, s40, v12 // 000000001AF0: D72C000E 00021828
|
|
v_add_co_u32 v73, vcc_lo, v7, v14 // 000000001AF8: D7006A49 00021D07
|
|
v_add_nc_u32_e32 v73, 8, v73 // 000000001B00: 4A929288
|
|
v_lshlrev_b32_e32 v73, 1, v73 // 000000001B04: 30929281
|
|
v_mul_lo_u32 v14, s40, v12 // 000000001B08: D72C000E 00021828
|
|
v_add_co_u32 v74, vcc_lo, v8, v14 // 000000001B10: D7006A4A 00021D08
|
|
v_add_nc_u32_e32 v74, 8, v74 // 000000001B18: 4A949488
|
|
v_lshlrev_b32_e32 v74, 1, v74 // 000000001B1C: 30949481
|
|
v_mul_lo_u32 v6, s42, v9 // 000000001B20: D72C0006 0002122A
|
|
v_add_co_u32 v75, vcc_lo, v13, v6 // 000000001B28: D7006A4B 00020D0D
|
|
v_add_nc_u32_e32 v75, 8, v75 // 000000001B30: 4A969688
|
|
v_lshlrev_b32_e32 v75, 1, v75 // 000000001B34: 30969681
|
|
v_mul_lo_u32 v6, s42, v10 // 000000001B38: D72C0006 0002142A
|
|
v_add_co_u32 v76, vcc_lo, v13, v6 // 000000001B40: D7006A4C 00020D0D
|
|
v_add_nc_u32_e32 v76, 8, v76 // 000000001B48: 4A989888
|
|
v_lshlrev_b32_e32 v76, 1, v76 // 000000001B4C: 30989881
|
|
v_mul_lo_u32 v6, s42, v11 // 000000001B50: D72C0006 0002162A
|
|
v_add_co_u32 v77, vcc_lo, v13, v6 // 000000001B58: D7006A4D 00020D0D
|
|
v_add_nc_u32_e32 v77, 8, v77 // 000000001B60: 4A9A9A88
|
|
v_lshlrev_b32_e32 v77, 1, v77 // 000000001B64: 309A9A81
|
|
s_mul_hi_u32 s69, s2, 0x60 // 000000001B68: 96C5FF02 00000060
|
|
s_mul_i32 s68, s2, 0x60 // 000000001B70: 9644FF02 00000060
|
|
s_mul_hi_u32 s67, 32, s6 // 000000001B78: 96C306A0
|
|
s_mul_i32 s66, 32, s6 // 000000001B7C: 964206A0
|
|
s_mul_hi_u32 s67, s66, s40 // 000000001B80: 96C32842
|
|
s_mul_i32 s66, s66, s40 // 000000001B84: 96422842
|
|
s_add_u32 s68, s68, s66 // 000000001B88: 80444244
|
|
s_addc_u32 s69, s69, s67 // 000000001B8C: 82454345
|
|
s_mov_b64 s[56:57], 1 // 000000001B90: BEB80181
|
|
s_sub_u32 s66, s24, 1 // 000000001B94: 80C28118
|
|
s_mul_hi_u32 s67, 1, s66 // 000000001B98: 96C34281
|
|
s_mul_i32 s66, 1, s66 // 000000001B9C: 96424281
|
|
s_add_u32 s56, s56, s66 // 000000001BA0: 80384238
|
|
s_addc_u32 s57, s57, s67 // 000000001BA4: 82394339
|
|
s_sub_u32 s66, s27, 1 // 000000001BA8: 80C2811B
|
|
s_mul_hi_u32 s67, s40, s66 // 000000001BAC: 96C34228
|
|
s_mul_i32 s66, s40, s66 // 000000001BB0: 96424228
|
|
s_add_u32 s56, s56, s66 // 000000001BB4: 80384238
|
|
s_addc_u32 s57, s57, s67 // 000000001BB8: 82394339
|
|
s_sub_u32 s56, s56, s68 // 000000001BBC: 80B84438
|
|
s_subb_u32 s57, s57, s69 // 000000001BC0: 82B94539
|
|
s_lshl_b64 s[56:57], s[56:57], 1 // 000000001BC4: 84B88138
|
|
s_add_u32 s56, s56, 16 // 000000001BC8: 80389038
|
|
s_addc_u32 s57, s57, 0 // 000000001BCC: 82398039
|
|
s_cmp_eq_u32 s57, 0 // 000000001BD0: BF068039
|
|
s_cselect_b32 s50, s56, -1 // 000000001BD4: 9832C138
|
|
s_mul_hi_u32 s67, s41, s4 // 000000001BD8: 96C30429
|
|
s_mul_i32 s66, s41, s4 // 000000001BDC: 96420429
|
|
s_add_u32 s68, s68, s66 // 000000001BE0: 80444244
|
|
s_addc_u32 s69, s69, s67 // 000000001BE4: 82454345
|
|
s_lshl_b64 s[68:69], s[68:69], 1 // 000000001BE8: 84C48144
|
|
s_add_u32 s48, s32, s68 // 000000001BEC: 80304420
|
|
s_addc_u32 s49, s33, s69 // 000000001BF0: 82314521
|
|
s_mov_b32 s51, 0x31004000 // 000000001BF4: BEB300FF 31004000
|
|
s_mul_hi_u32 s69, s3, 0x60 // 000000001BFC: 96C5FF03 00000060
|
|
s_mul_i32 s68, s3, 0x60 // 000000001C04: 9644FF03 00000060
|
|
s_mul_hi_u32 s69, s68, s42 // 000000001C0C: 96C52A44
|
|
s_mul_i32 s68, s68, s42 // 000000001C10: 96442A44
|
|
s_mul_hi_u32 s67, 32, s6 // 000000001C14: 96C306A0
|
|
s_mul_i32 s66, 32, s6 // 000000001C18: 964206A0
|
|
s_add_u32 s68, s68, s66 // 000000001C1C: 80444244
|
|
s_addc_u32 s69, s69, s67 // 000000001C20: 82454345
|
|
s_mov_b64 s[58:59], 1 // 000000001C24: BEBA0181
|
|
s_sub_u32 s66, s27, 1 // 000000001C28: 80C2811B
|
|
s_mul_hi_u32 s67, 1, s66 // 000000001C2C: 96C34281
|
|
s_mul_i32 s66, 1, s66 // 000000001C30: 96424281
|
|
s_add_u32 s58, s58, s66 // 000000001C34: 803A423A
|
|
s_addc_u32 s59, s59, s67 // 000000001C38: 823B433B
|
|
s_sub_u32 s66, s25, 1 // 000000001C3C: 80C28119
|
|
s_mul_hi_u32 s67, s42, s66 // 000000001C40: 96C3422A
|
|
s_mul_i32 s66, s42, s66 // 000000001C44: 9642422A
|
|
s_add_u32 s58, s58, s66 // 000000001C48: 803A423A
|
|
s_addc_u32 s59, s59, s67 // 000000001C4C: 823B433B
|
|
s_sub_u32 s58, s58, s68 // 000000001C50: 80BA443A
|
|
s_subb_u32 s59, s59, s69 // 000000001C54: 82BB453B
|
|
s_lshl_b64 s[58:59], s[58:59], 1 // 000000001C58: 84BA813A
|
|
s_add_u32 s58, s58, 16 // 000000001C5C: 803A903A
|
|
s_addc_u32 s59, s59, 0 // 000000001C60: 823B803B
|
|
s_cmp_eq_u32 s59, 0 // 000000001C64: BF06803B
|
|
s_cselect_b32 s54, s58, -1 // 000000001C68: 9836C13A
|
|
s_mul_hi_u32 s67, s43, s4 // 000000001C6C: 96C3042B
|
|
s_mul_i32 s66, s43, s4 // 000000001C70: 9642042B
|
|
s_add_u32 s68, s68, s66 // 000000001C74: 80444244
|
|
s_addc_u32 s69, s69, s67 // 000000001C78: 82454345
|
|
s_lshl_b64 s[68:69], s[68:69], 1 // 000000001C7C: 84C48144
|
|
s_add_u32 s52, s34, s68 // 000000001C80: 80344422
|
|
s_addc_u32 s53, s35, s69 // 000000001C84: 82354523
|
|
s_mov_b32 s55, 0x31004000 // 000000001C88: BEB700FF 31004000
|
|
s_and_b32 s67, s46, 0x3fff // 000000001C90: 8B43FF2E 00003FFF
|
|
s_mul_i32 s67, s67, 64 // 000000001C98: 9643C043
|
|
s_and_b32 s66, s46, 0x8000 // 000000001C9C: 8B42FF2E 00008000
|
|
s_cmov_b32 s67, 64 // 000000001CA4: BEC302C0
|
|
s_mul_i32 s64, s67, s40 // 000000001CA8: 96402843
|
|
s_and_b32 s67, s46, 0x3fff // 000000001CAC: 8B43FF2E 00003FFF
|
|
s_mul_i32 s67, s67, 64 // 000000001CB4: 9643C043
|
|
s_and_b32 s66, s46, 0x8000 // 000000001CB8: 8B42FF2E 00008000
|
|
s_cselect_b32 s65, 64, s67 // 000000001CC0: 984143C0
|
|
s_lshr_b32 s12, s27, 5 // 000000001CC4: 850C851B
|
|
s_mov_b32 s13, s12 // 000000001CC8: BE8D000C
|
|
s_and_b32 s68, s10, 0x1f00 // 000000001CCC: 8B44FF0A 00001F00
|
|
s_lshr_b32 s68, s68, 8 // 000000001CD4: 85448844
|
|
s_and_b32 s69, s10, 0xe000 // 000000001CD8: 8B45FF0A 0000E000
|
|
s_and_b32 s10, s10, 0xff // 000000001CE0: 8B0AFF0A 000000FF
|
|
s_mov_b32 s66, s10 // 000000001CE8: BEC2000A
|
|
s_lshl_b32 s67, s66, s68 // 000000001CEC: 84434442
|
|
s_cmp_ge_u32 s13, s67 // 000000001CF0: BF09430D
|
|
s_sub_u32 s67, s66, 1 // 000000001CF4: 80C38142
|
|
s_cmp_ge_u32 s66, 1 // 000000001CF8: BF098142
|
|
s_cselect_b32 s47, s67, 0 // 000000001CFC: 982F8043
|
|
s_cmp_eq_u32 s69, 0 // 000000001D00: BF068045
|
|
s_and_b32 s47, s47, s66 // 000000001D04: 8B2F422F
|
|
s_lshl_b32 s47, s47, s68 // 000000001D08: 842F442F
|
|
s_mul_hi_i32 s67, s47, s64 // 000000001D0C: 9743402F
|
|
s_mul_i32 s66, s47, s64 // 000000001D10: 9642402F
|
|
s_mul_hi_i32 s61, s12, s64 // 000000001D14: 973D400C
|
|
s_mul_i32 s60, s12, s64 // 000000001D18: 963C400C
|
|
s_sub_u32 s60, s64, s60 // 000000001D1C: 80BC3C40
|
|
s_subb_u32 s61, 0, s61 // 000000001D20: 82BD3D80
|
|
s_add_u32 s48, s48, s66 // 000000001D24: 80304230
|
|
s_addc_u32 s49, s49, s67 // 000000001D28: 82314331
|
|
s_sub_u32 s56, s56, s66 // 000000001D2C: 80B84238
|
|
s_subb_u32 s57, s57, s67 // 000000001D30: 82B94339
|
|
s_cmp_eq_u32 s57, 0 // 000000001D34: BF068039
|
|
s_cselect_b32 s50, s56, -1 // 000000001D38: 9832C138
|
|
s_mul_hi_i32 s67, s47, s65 // 000000001D3C: 9743412F
|
|
s_mul_i32 s66, s47, s65 // 000000001D40: 9642412F
|
|
s_mul_hi_i32 s63, s12, s65 // 000000001D44: 973F410C
|
|
s_mul_i32 s62, s12, s65 // 000000001D48: 963E410C
|
|
s_sub_u32 s62, s65, s62 // 000000001D4C: 80BE3E41
|
|
s_subb_u32 s63, 0, s63 // 000000001D50: 82BF3F80
|
|
s_add_u32 s52, s52, s66 // 000000001D54: 80344234
|
|
s_addc_u32 s53, s53, s67 // 000000001D58: 82354335
|
|
s_sub_u32 s58, s58, s66 // 000000001D5C: 80BA423A
|
|
s_subb_u32 s59, s59, s67 // 000000001D60: 82BB433B
|
|
s_cmp_eq_u32 s59, 0 // 000000001D64: BF06803B
|
|
s_cselect_b32 s54, s58, -1 // 000000001D68: 9836C13A
|
|
s_add_u32 s47, s47, 2 // 000000001D6C: 802F822F
|
|
buffer_load_b128 v[230:233], v72, s[48:51], 0 offen // 000000001D70: E05C0000 804CE648
|
|
buffer_load_b128 v[234:237], v73, s[48:51], 0 offen // 000000001D78: E05C0000 804CEA49
|
|
buffer_load_b128 v[238:241], v74, s[48:51], 0 offen // 000000001D80: E05C0000 804CEE4A
|
|
buffer_load_b128 v[242:245], v75, s[52:55], 0 offen // 000000001D88: E05C0000 804DF24B
|
|
buffer_load_b128 v[246:249], v76, s[52:55], 0 offen // 000000001D90: E05C0000 804DF64C
|
|
buffer_load_b128 v[250:253], v77, s[52:55], 0 offen // 000000001D98: E05C0000 804DFA4D
|
|
s_add_u32 s68, s12, 1 // 000000001DA0: 8044810C
|
|
s_cmp_eq_u32 s47, s68 // 000000001DA4: BF06442F
|
|
s_cselect_b32 s66, s60, s64 // 000000001DA8: 9842403C
|
|
s_cselect_b32 s67, s61, 0 // 000000001DAC: 9843803D
|
|
s_add_u32 s48, s48, s66 // 000000001DB0: 80304230
|
|
s_addc_u32 s49, s49, s67 // 000000001DB4: 82314331
|
|
s_sub_u32 s56, s56, s66 // 000000001DB8: 80B84238
|
|
s_subb_u32 s57, s57, s67 // 000000001DBC: 82B94339
|
|
s_cmp_eq_u32 s57, 0 // 000000001DC0: BF068039
|
|
s_cselect_b32 s50, s56, -1 // 000000001DC4: 9832C138
|
|
s_add_u32 s68, s12, 1 // 000000001DC8: 8044810C
|
|
s_cmp_eq_u32 s47, s68 // 000000001DCC: BF06442F
|
|
s_cselect_b32 s66, s62, s65 // 000000001DD0: 9842413E
|
|
s_cselect_b32 s67, s63, 0 // 000000001DD4: 9843803F
|
|
s_add_u32 s52, s52, s66 // 000000001DD8: 80344234
|
|
s_addc_u32 s53, s53, s67 // 000000001DDC: 82354335
|
|
s_sub_u32 s58, s58, s66 // 000000001DE0: 80BA423A
|
|
s_subb_u32 s59, s59, s67 // 000000001DE4: 82BB433B
|
|
s_cmp_eq_u32 s59, 0 // 000000001DE8: BF06803B
|
|
s_cselect_b32 s54, s58, -1 // 000000001DEC: 9836C13A
|
|
s_mov_b64 s[16:17], s[28:29] // 000000001DF0: BE90011C
|
|
s_mov_b32 s18, 0x80000000 // 000000001DF4: BE9200FF 80000000
|
|
s_mov_b32 s19, 0x31004000 // 000000001DFC: BE9300FF 31004000
|
|
s_mov_b64 s[20:21], s[30:31] // 000000001E04: BE94011E
|
|
s_mov_b32 s22, 0x80000000 // 000000001E08: BE9600FF 80000000
|
|
s_mov_b32 s23, 0x31004000 // 000000001E10: BE9700FF 31004000
|
|
s_mul_i32 s68, 0x60, s3 // 000000001E18: 964403FF 00000060
|
|
s_mul_hi_u32 s67, s68, s38 // 000000001E20: 96C32644
|
|
s_mul_i32 s66, s68, s38 // 000000001E24: 96422644
|
|
s_lshl_b64 s[66:67], s[66:67], s8 // 000000001E28: 84C20842
|
|
s_add_u32 s20, s30, s66 // 000000001E2C: 8014421E
|
|
s_addc_u32 s21, s31, s67 // 000000001E30: 8215431F
|
|
s_mul_hi_u32 s67, s68, s36 // 000000001E34: 96C32444
|
|
s_mul_i32 s66, s68, s36 // 000000001E38: 96422444
|
|
s_lshl_b64 s[66:67], s[66:67], s9 // 000000001E3C: 84C20942
|
|
s_add_u32 s16, s28, s66 // 000000001E40: 8010421C
|
|
s_addc_u32 s17, s29, s67 // 000000001E44: 8211431D
|
|
s_mul_hi_u32 s67, s4, s39 // 000000001E48: 96C32704
|
|
s_mul_i32 s66, s4, s39 // 000000001E4C: 96422704
|
|
s_lshl_b64 s[66:67], s[66:67], s8 // 000000001E50: 84C20842
|
|
s_add_u32 s20, s20, s66 // 000000001E54: 80144214
|
|
s_addc_u32 s21, s21, s67 // 000000001E58: 82154315
|
|
s_mul_hi_u32 s67, s4, s37 // 000000001E5C: 96C32504
|
|
s_mul_i32 s66, s4, s37 // 000000001E60: 96422504
|
|
s_lshl_b64 s[66:67], s[66:67], s9 // 000000001E64: 84C20942
|
|
s_add_u32 s16, s16, s66 // 000000001E68: 80104210
|
|
s_addc_u32 s17, s17, s67 // 000000001E6C: 82114311
|
|
v_mov_b32_e32 v0, 0 // 000000001E70: 7E000280
|
|
v_mov_b32_e32 v1, 0 // 000000001E74: 7E020280
|
|
v_mov_b32_e32 v2, 0 // 000000001E78: 7E040280
|
|
v_mov_b32_e32 v3, 0 // 000000001E7C: 7E060280
|
|
v_mov_b32_e32 v4, 0 // 000000001E80: 7E080280
|
|
v_mov_b32_e32 v5, 0 // 000000001E84: 7E0A0280
|
|
v_mov_b32_e32 v6, 0 // 000000001E88: 7E0C0280
|
|
v_mov_b32_e32 v7, 0 // 000000001E8C: 7E0E0280
|
|
v_mov_b32_e32 v8, 0 // 000000001E90: 7E100280
|
|
v_mov_b32_e32 v9, 0 // 000000001E94: 7E120280
|
|
v_mov_b32_e32 v10, 0 // 000000001E98: 7E140280
|
|
v_mov_b32_e32 v11, 0 // 000000001E9C: 7E160280
|
|
v_mov_b32_e32 v12, 0 // 000000001EA0: 7E180280
|
|
v_mov_b32_e32 v13, 0 // 000000001EA4: 7E1A0280
|
|
v_mov_b32_e32 v14, 0 // 000000001EA8: 7E1C0280
|
|
v_mov_b32_e32 v15, 0 // 000000001EAC: 7E1E0280
|
|
v_mov_b32_e32 v16, 0 // 000000001EB0: 7E200280
|
|
v_mov_b32_e32 v17, 0 // 000000001EB4: 7E220280
|
|
v_mov_b32_e32 v18, 0 // 000000001EB8: 7E240280
|
|
v_mov_b32_e32 v19, 0 // 000000001EBC: 7E260280
|
|
v_mov_b32_e32 v20, 0 // 000000001EC0: 7E280280
|
|
v_mov_b32_e32 v21, 0 // 000000001EC4: 7E2A0280
|
|
v_mov_b32_e32 v22, 0 // 000000001EC8: 7E2C0280
|
|
v_mov_b32_e32 v23, 0 // 000000001ECC: 7E2E0280
|
|
v_mov_b32_e32 v24, 0 // 000000001ED0: 7E300280
|
|
v_mov_b32_e32 v25, 0 // 000000001ED4: 7E320280
|
|
v_mov_b32_e32 v26, 0 // 000000001ED8: 7E340280
|
|
v_mov_b32_e32 v27, 0 // 000000001EDC: 7E360280
|
|
v_mov_b32_e32 v28, 0 // 000000001EE0: 7E380280
|
|
v_mov_b32_e32 v29, 0 // 000000001EE4: 7E3A0280
|
|
v_mov_b32_e32 v30, 0 // 000000001EE8: 7E3C0280
|
|
v_mov_b32_e32 v31, 0 // 000000001EEC: 7E3E0280
|
|
v_mov_b32_e32 v32, 0 // 000000001EF0: 7E400280
|
|
v_mov_b32_e32 v33, 0 // 000000001EF4: 7E420280
|
|
v_mov_b32_e32 v34, 0 // 000000001EF8: 7E440280
|
|
v_mov_b32_e32 v35, 0 // 000000001EFC: 7E460280
|
|
v_mov_b32_e32 v36, 0 // 000000001F00: 7E480280
|
|
v_mov_b32_e32 v37, 0 // 000000001F04: 7E4A0280
|
|
v_mov_b32_e32 v38, 0 // 000000001F08: 7E4C0280
|
|
v_mov_b32_e32 v39, 0 // 000000001F0C: 7E4E0280
|
|
v_mov_b32_e32 v40, 0 // 000000001F10: 7E500280
|
|
v_mov_b32_e32 v41, 0 // 000000001F14: 7E520280
|
|
v_mov_b32_e32 v42, 0 // 000000001F18: 7E540280
|
|
v_mov_b32_e32 v43, 0 // 000000001F1C: 7E560280
|
|
v_mov_b32_e32 v44, 0 // 000000001F20: 7E580280
|
|
v_mov_b32_e32 v45, 0 // 000000001F24: 7E5A0280
|
|
v_mov_b32_e32 v46, 0 // 000000001F28: 7E5C0280
|
|
v_mov_b32_e32 v47, 0 // 000000001F2C: 7E5E0280
|
|
v_mov_b32_e32 v48, 0 // 000000001F30: 7E600280
|
|
v_mov_b32_e32 v49, 0 // 000000001F34: 7E620280
|
|
v_mov_b32_e32 v50, 0 // 000000001F38: 7E640280
|
|
v_mov_b32_e32 v51, 0 // 000000001F3C: 7E660280
|
|
v_mov_b32_e32 v52, 0 // 000000001F40: 7E680280
|
|
v_mov_b32_e32 v53, 0 // 000000001F44: 7E6A0280
|
|
v_mov_b32_e32 v54, 0 // 000000001F48: 7E6C0280
|
|
v_mov_b32_e32 v55, 0 // 000000001F4C: 7E6E0280
|
|
v_mov_b32_e32 v56, 0 // 000000001F50: 7E700280
|
|
v_mov_b32_e32 v57, 0 // 000000001F54: 7E720280
|
|
v_mov_b32_e32 v58, 0 // 000000001F58: 7E740280
|
|
v_mov_b32_e32 v59, 0 // 000000001F5C: 7E760280
|
|
v_mov_b32_e32 v60, 0 // 000000001F60: 7E780280
|
|
v_mov_b32_e32 v61, 0 // 000000001F64: 7E7A0280
|
|
v_mov_b32_e32 v62, 0 // 000000001F68: 7E7C0280
|
|
v_mov_b32_e32 v63, 0 // 000000001F6C: 7E7E0280
|
|
v_mov_b32_e32 v64, 0 // 000000001F70: 7E800280
|
|
v_mov_b32_e32 v65, 0 // 000000001F74: 7E820280
|
|
v_mov_b32_e32 v66, 0 // 000000001F78: 7E840280
|
|
v_mov_b32_e32 v67, 0 // 000000001F7C: 7E860280
|
|
v_mov_b32_e32 v68, 0 // 000000001F80: 7E880280
|
|
v_mov_b32_e32 v69, 0 // 000000001F84: 7E8A0280
|
|
v_mov_b32_e32 v70, 0 // 000000001F88: 7E8C0280
|
|
v_mov_b32_e32 v71, 0 // 000000001F8C: 7E8E0280
|
|
s_cmp_eq_u32 s12, 0 // 000000001F90: BF06800C
|
|
s_waitcnt vmcnt(0) // 000000001F94: BF8903F7
|
|
ds_store_b128 v78, v[230:233] // 000000001F98: DB7C0000 0000E64E
|
|
ds_store_b128 v78, v[234:237] offset:64 // 000000001FA0: DB7C0040 0000EA4E
|
|
ds_store_b128 v78, v[238:241] offset:128 // 000000001FA8: DB7C0080 0000EE4E
|
|
ds_store_b128 v79, v[242:245] // 000000001FB0: DB7C0000 0000F24F
|
|
ds_store_b128 v79, v[246:249] offset:2560 // 000000001FB8: DB7C0A00 0000F64F
|
|
ds_store_b128 v79, v[250:253] offset:5120 // 000000001FC0: DB7C1400 0000FA4F
|
|
v_xor_b32_e32 v78, 0x4000, v78 // 000000001FC8: 3A9C9CFF 00004000
|
|
v_xor_b32_e32 v79, 0x4000, v79 // 000000001FD0: 3A9E9EFF 00004000
|
|
buffer_load_b128 v[230:233], v72, s[48:51], 0 offen // 000000001FD8: E05C0000 804CE648
|
|
buffer_load_b128 v[234:237], v73, s[48:51], 0 offen // 000000001FE0: E05C0000 804CEA49
|
|
buffer_load_b128 v[238:241], v74, s[48:51], 0 offen // 000000001FE8: E05C0000 804CEE4A
|
|
buffer_load_b128 v[242:245], v75, s[52:55], 0 offen // 000000001FF0: E05C0000 804DF24B
|
|
buffer_load_b128 v[246:249], v76, s[52:55], 0 offen // 000000001FF8: E05C0000 804DF64C
|
|
buffer_load_b128 v[250:253], v77, s[52:55], 0 offen // 000000002000: E05C0000 804DFA4D
|
|
s_waitcnt lgkmcnt(0) // 000000002008: BF89FC07
|
|
s_waitcnt lgkmcnt(0) // 00000000200C: BF89FC07
|
|
s_barrier // 000000002010: BFBD0000
|
|
ds_load_u16 v84, v80 // 000000002014: D8F00000 54000050
|
|
ds_load_u16_d16_hi v84, v80 offset:192 // 00000000201C: DA9C00C0 54000050
|
|
ds_load_u16 v85, v80 offset:384 // 000000002024: D8F00180 55000050
|
|
ds_load_u16_d16_hi v85, v80 offset:576 // 00000000202C: DA9C0240 55000050
|
|
ds_load_u16 v86, v80 offset:768 // 000000002034: D8F00300 56000050
|
|
ds_load_u16_d16_hi v86, v80 offset:960 // 00000000203C: DA9C03C0 56000050
|
|
ds_load_u16 v87, v80 offset:1152 // 000000002044: D8F00480 57000050
|
|
ds_load_u16_d16_hi v87, v80 offset:1344 // 00000000204C: DA9C0540 57000050
|
|
ds_load_u16 v88, v80 offset:1536 // 000000002054: D8F00600 58000050
|
|
ds_load_u16_d16_hi v88, v80 offset:1728 // 00000000205C: DA9C06C0 58000050
|
|
ds_load_u16 v89, v80 offset:1920 // 000000002064: D8F00780 59000050
|
|
ds_load_u16_d16_hi v89, v80 offset:2112 // 00000000206C: DA9C0840 59000050
|
|
ds_load_u16 v90, v80 offset:2304 // 000000002074: D8F00900 5A000050
|
|
ds_load_u16_d16_hi v90, v80 offset:2496 // 00000000207C: DA9C09C0 5A000050
|
|
ds_load_u16 v91, v80 offset:2688 // 000000002084: D8F00A80 5B000050
|
|
ds_load_u16_d16_hi v91, v80 offset:2880 // 00000000208C: DA9C0B40 5B000050
|
|
ds_load_u16 v92, v80 offset:64 // 000000002094: D8F00040 5C000050
|
|
ds_load_u16_d16_hi v92, v80 offset:256 // 00000000209C: DA9C0100 5C000050
|
|
ds_load_u16 v93, v80 offset:448 // 0000000020A4: D8F001C0 5D000050
|
|
ds_load_u16_d16_hi v93, v80 offset:640 // 0000000020AC: DA9C0280 5D000050
|
|
ds_load_u16 v94, v80 offset:832 // 0000000020B4: D8F00340 5E000050
|
|
ds_load_u16_d16_hi v94, v80 offset:1024 // 0000000020BC: DA9C0400 5E000050
|
|
ds_load_u16 v95, v80 offset:1216 // 0000000020C4: D8F004C0 5F000050
|
|
ds_load_u16_d16_hi v95, v80 offset:1408 // 0000000020CC: DA9C0580 5F000050
|
|
ds_load_u16 v96, v80 offset:1600 // 0000000020D4: D8F00640 60000050
|
|
ds_load_u16_d16_hi v96, v80 offset:1792 // 0000000020DC: DA9C0700 60000050
|
|
ds_load_u16 v97, v80 offset:1984 // 0000000020E4: D8F007C0 61000050
|
|
ds_load_u16_d16_hi v97, v80 offset:2176 // 0000000020EC: DA9C0880 61000050
|
|
ds_load_u16 v98, v80 offset:2368 // 0000000020F4: D8F00940 62000050
|
|
ds_load_u16_d16_hi v98, v80 offset:2560 // 0000000020FC: DA9C0A00 62000050
|
|
ds_load_u16 v99, v80 offset:2752 // 000000002104: D8F00AC0 63000050
|
|
ds_load_u16_d16_hi v99, v80 offset:2944 // 00000000210C: DA9C0B80 63000050
|
|
ds_load_u16 v100, v80 offset:128 // 000000002114: D8F00080 64000050
|
|
ds_load_u16_d16_hi v100, v80 offset:320 // 00000000211C: DA9C0140 64000050
|
|
ds_load_u16 v101, v80 offset:512 // 000000002124: D8F00200 65000050
|
|
ds_load_u16_d16_hi v101, v80 offset:704 // 00000000212C: DA9C02C0 65000050
|
|
ds_load_u16 v102, v80 offset:896 // 000000002134: D8F00380 66000050
|
|
ds_load_u16_d16_hi v102, v80 offset:1088 // 00000000213C: DA9C0440 66000050
|
|
ds_load_u16 v103, v80 offset:1280 // 000000002144: D8F00500 67000050
|
|
ds_load_u16_d16_hi v103, v80 offset:1472 // 00000000214C: DA9C05C0 67000050
|
|
ds_load_u16 v104, v80 offset:1664 // 000000002154: D8F00680 68000050
|
|
ds_load_u16_d16_hi v104, v80 offset:1856 // 00000000215C: DA9C0740 68000050
|
|
ds_load_u16 v105, v80 offset:2048 // 000000002164: D8F00800 69000050
|
|
ds_load_u16_d16_hi v105, v80 offset:2240 // 00000000216C: DA9C08C0 69000050
|
|
ds_load_u16 v106, v80 offset:2432 // 000000002174: D8F00980 6A000050
|
|
ds_load_u16_d16_hi v106, v80 offset:2624 // 00000000217C: DA9C0A40 6A000050
|
|
ds_load_u16 v107, v80 offset:2816 // 000000002184: D8F00B00 6B000050
|
|
ds_load_u16_d16_hi v107, v80 offset:3008 // 00000000218C: DA9C0BC0 6B000050
|
|
ds_load_b128 v[181:184], v81 // 000000002194: DBFC0000 B5000051
|
|
ds_load_b128 v[185:188], v81 offset:16 // 00000000219C: DBFC0010 B9000051
|
|
ds_load_b128 v[189:192], v81 offset:2560 // 0000000021A4: DBFC0A00 BD000051
|
|
ds_load_b128 v[193:196], v81 offset:2576 // 0000000021AC: DBFC0A10 C1000051
|
|
ds_load_b128 v[197:200], v81 offset:5120 // 0000000021B4: DBFC1400 C5000051
|
|
ds_load_b128 v[201:204], v81 offset:5136 // 0000000021BC: DBFC1410 C9000051
|
|
|
|
main_loop: // 00000000000021c4
|
|
s_waitcnt lgkmcnt(4) // 0000000021C4: BF89FC47
|
|
v_wmma_f32_16x16x16_f16 v[0:7], v[181:188], v[84:91], v[0:7]// 0000000021C8: CC404000 1C02A9B5
|
|
ds_load_u16 v108, v80 offset:3104 // 0000000021D0: D8F00C20 6C000050
|
|
ds_load_u16_d16_hi v108, v80 offset:3296 // 0000000021D8: DA9C0CE0 6C000050
|
|
ds_load_u16 v109, v80 offset:3488 // 0000000021E0: D8F00DA0 6D000050
|
|
ds_load_u16_d16_hi v109, v80 offset:3680 // 0000000021E8: DA9C0E60 6D000050
|
|
ds_load_u16 v110, v80 offset:3872 // 0000000021F0: D8F00F20 6E000050
|
|
ds_load_u16_d16_hi v110, v80 offset:4064 // 0000000021F8: DA9C0FE0 6E000050
|
|
ds_load_u16 v111, v80 offset:4256 // 000000002200: D8F010A0 6F000050
|
|
s_cmp_eq_u32 s12, s47 // 000000002208: BF062F0C
|
|
s_cselect_b32 s66, s60, s64 // 00000000220C: 9842403C
|
|
s_cselect_b32 s67, s61, 0 // 000000002210: 9843803D
|
|
v_wmma_f32_16x16x16_f16 v[8:15], v[181:188], v[92:99], v[8:15]// 000000002214: CC404008 1C22B9B5
|
|
ds_load_u16_d16_hi v111, v80 offset:4448 // 00000000221C: DA9C1160 6F000050
|
|
ds_load_u16 v112, v80 offset:4640 // 000000002224: D8F01220 70000050
|
|
ds_load_u16_d16_hi v112, v80 offset:4832 // 00000000222C: DA9C12E0 70000050
|
|
ds_load_u16 v113, v80 offset:5024 // 000000002234: D8F013A0 71000050
|
|
ds_load_u16_d16_hi v113, v80 offset:5216 // 00000000223C: DA9C1460 71000050
|
|
ds_load_u16 v114, v80 offset:5408 // 000000002244: D8F01520 72000050
|
|
ds_load_u16_d16_hi v114, v80 offset:5600 // 00000000224C: DA9C15E0 72000050
|
|
s_add_u32 s48, s48, s66 // 000000002254: 80304230
|
|
s_addc_u32 s49, s49, s67 // 000000002258: 82314331
|
|
s_sub_u32 s56, s56, s66 // 00000000225C: 80B84238
|
|
v_wmma_f32_16x16x16_f16 v[16:23], v[181:188], v[100:107], v[16:23]// 000000002260: CC404010 1C42C9B5
|
|
ds_load_u16 v115, v80 offset:5792 // 000000002268: D8F016A0 73000050
|
|
ds_load_u16_d16_hi v115, v80 offset:5984 // 000000002270: DA9C1760 73000050
|
|
ds_load_b128 v[205:208], v81 offset:32 // 000000002278: DBFC0020 CD000051
|
|
ds_load_b128 v[209:212], v81 offset:48 // 000000002280: DBFC0030 D1000051
|
|
ds_load_u16 v116, v80 offset:3168 // 000000002288: D8F00C60 74000050
|
|
ds_load_u16_d16_hi v116, v80 offset:3360 // 000000002290: DA9C0D20 74000050
|
|
ds_load_u16 v117, v80 offset:3552 // 000000002298: D8F00DE0 75000050
|
|
s_subb_u32 s57, s57, s67 // 0000000022A0: 82B94339
|
|
s_cmp_eq_u32 s57, 0 // 0000000022A4: BF068039
|
|
s_cselect_b32 s50, s56, -1 // 0000000022A8: 9832C138
|
|
s_waitcnt lgkmcnt(21) // 0000000022AC: BF89FD57
|
|
v_wmma_f32_16x16x16_f16 v[24:31], v[189:196], v[84:91], v[24:31]// 0000000022B0: CC404018 1C62A9BD
|
|
ds_load_u16_d16_hi v117, v80 offset:3744 // 0000000022B8: DA9C0EA0 75000050
|
|
ds_load_u16 v118, v80 offset:3936 // 0000000022C0: D8F00F60 76000050
|
|
ds_load_u16_d16_hi v118, v80 offset:4128 // 0000000022C8: DA9C1020 76000050
|
|
ds_load_u16 v119, v80 offset:4320 // 0000000022D0: D8F010E0 77000050
|
|
ds_load_u16_d16_hi v119, v80 offset:4512 // 0000000022D8: DA9C11A0 77000050
|
|
ds_load_u16 v120, v80 offset:4704 // 0000000022E0: D8F01260 78000050
|
|
ds_load_u16_d16_hi v120, v80 offset:4896 // 0000000022E8: DA9C1320 78000050
|
|
s_cmp_eq_u32 s12, s47 // 0000000022F0: BF062F0C
|
|
s_cselect_b32 s66, s62, s65 // 0000000022F4: 9842413E
|
|
s_cselect_b32 s67, s63, 0 // 0000000022F8: 9843803F
|
|
v_wmma_f32_16x16x16_f16 v[32:39], v[189:196], v[92:99], v[32:39]// 0000000022FC: CC404020 1C82B9BD
|
|
ds_load_u16 v121, v80 offset:5088 // 000000002304: D8F013E0 79000050
|
|
ds_load_u16_d16_hi v121, v80 offset:5280 // 00000000230C: DA9C14A0 79000050
|
|
ds_load_u16 v122, v80 offset:5472 // 000000002314: D8F01560 7A000050
|
|
ds_load_u16_d16_hi v122, v80 offset:5664 // 00000000231C: DA9C1620 7A000050
|
|
ds_load_u16 v123, v80 offset:5856 // 000000002324: D8F016E0 7B000050
|
|
ds_load_u16_d16_hi v123, v80 offset:6048 // 00000000232C: DA9C17A0 7B000050
|
|
ds_load_u16 v124, v80 offset:3232 // 000000002334: D8F00CA0 7C000050
|
|
s_add_u32 s52, s52, s66 // 00000000233C: 80344234
|
|
s_addc_u32 s53, s53, s67 // 000000002340: 82354335
|
|
s_sub_u32 s58, s58, s66 // 000000002344: 80BA423A
|
|
v_wmma_f32_16x16x16_f16 v[40:47], v[189:196], v[100:107], v[40:47]// 000000002348: CC404028 1CA2C9BD
|
|
ds_load_u16_d16_hi v124, v80 offset:3424 // 000000002350: DA9C0D60 7C000050
|
|
ds_load_u16 v125, v80 offset:3616 // 000000002358: D8F00E20 7D000050
|
|
ds_load_u16_d16_hi v125, v80 offset:3808 // 000000002360: DA9C0EE0 7D000050
|
|
ds_load_u16 v126, v80 offset:4000 // 000000002368: D8F00FA0 7E000050
|
|
ds_load_u16_d16_hi v126, v80 offset:4192 // 000000002370: DA9C1060 7E000050
|
|
ds_load_u16 v127, v80 offset:4384 // 000000002378: D8F01120 7F000050
|
|
ds_load_u16_d16_hi v127, v80 offset:4576 // 000000002380: DA9C11E0 7F000050
|
|
s_subb_u32 s59, s59, s67 // 000000002388: 82BB433B
|
|
s_cmp_eq_u32 s59, 0 // 00000000238C: BF06803B
|
|
s_cselect_b32 s54, s58, -1 // 000000002390: 9836C13A
|
|
s_waitcnt vmcnt(5) // 000000002394: BF8917F7
|
|
ds_store_b128 v78, v[230:233] // 000000002398: DB7C0000 0000E64E
|
|
buffer_load_b128 v[230:233], v72, s[48:51], 0 offen // 0000000023A0: E05C0000 804CE648
|
|
s_waitcnt vmcnt(5) // 0000000023A8: BF8917F7
|
|
ds_store_b128 v78, v[234:237] offset:64 // 0000000023AC: DB7C0040 0000EA4E
|
|
buffer_load_b128 v[234:237], v73, s[48:51], 0 offen // 0000000023B4: E05C0000 804CEA49
|
|
s_waitcnt vmcnt(5) // 0000000023BC: BF8917F7
|
|
ds_store_b128 v78, v[238:241] offset:128 // 0000000023C0: DB7C0080 0000EE4E
|
|
buffer_load_b128 v[238:241], v74, s[48:51], 0 offen // 0000000023C8: E05C0000 804CEE4A
|
|
s_waitcnt vmcnt(5) // 0000000023D0: BF8917F7
|
|
ds_store_b128 v79, v[242:245] // 0000000023D4: DB7C0000 0000F24F
|
|
buffer_load_b128 v[242:245], v75, s[52:55], 0 offen // 0000000023DC: E05C0000 804DF24B
|
|
s_waitcnt vmcnt(5) // 0000000023E4: BF8917F7
|
|
ds_store_b128 v79, v[246:249] offset:2560 // 0000000023E8: DB7C0A00 0000F64F
|
|
buffer_load_b128 v[246:249], v76, s[52:55], 0 offen // 0000000023F0: E05C0000 804DF64C
|
|
s_waitcnt vmcnt(5) // 0000000023F8: BF8917F7
|
|
ds_store_b128 v79, v[250:253] offset:5120 // 0000000023FC: DB7C1400 0000FA4F
|
|
buffer_load_b128 v[250:253], v77, s[52:55], 0 offen // 000000002404: E05C0000 804DFA4D
|
|
v_xor_b32_e32 v78, 0x4000, v78 // 00000000240C: 3A9C9CFF 00004000
|
|
v_xor_b32_e32 v79, 0x4000, v79 // 000000002414: 3A9E9EFF 00004000
|
|
v_wmma_f32_16x16x16_f16 v[48:55], v[197:204], v[84:91], v[48:55]// 00000000241C: CC404030 1CC2A9C5
|
|
ds_load_u16 v128, v80 offset:4768 // 000000002424: D8F012A0 80000050
|
|
ds_load_u16_d16_hi v128, v80 offset:4960 // 00000000242C: DA9C1360 80000050
|
|
ds_load_u16 v129, v80 offset:5152 // 000000002434: D8F01420 81000050
|
|
ds_load_u16_d16_hi v129, v80 offset:5344 // 00000000243C: DA9C14E0 81000050
|
|
ds_load_u16 v130, v80 offset:5536 // 000000002444: D8F015A0 82000050
|
|
ds_load_u16_d16_hi v130, v80 offset:5728 // 00000000244C: DA9C1660 82000050
|
|
v_wmma_f32_16x16x16_f16 v[56:63], v[197:204], v[92:99], v[56:63]// 000000002454: CC404038 1CE2B9C5
|
|
ds_load_u16 v131, v80 offset:5920 // 00000000245C: D8F01720 83000050
|
|
ds_load_u16_d16_hi v131, v80 offset:6112 // 000000002464: DA9C17E0 83000050
|
|
ds_load_b128 v[213:216], v81 offset:2592 // 00000000246C: DBFC0A20 D5000051
|
|
ds_load_b128 v[217:220], v81 offset:2608 // 000000002474: DBFC0A30 D9000051
|
|
ds_load_b128 v[221:224], v81 offset:5152 // 00000000247C: DBFC1420 DD000051
|
|
ds_load_b128 v[225:228], v81 offset:5168 // 000000002484: DBFC1430 E1000051
|
|
v_xor_b32_e32 v80, 0x4000, v80 // 00000000248C: 3AA0A0FF 00004000
|
|
v_xor_b32_e32 v81, 0x4000, v81 // 000000002494: 3AA2A2FF 00004000
|
|
v_wmma_f32_16x16x16_f16 v[64:71], v[197:204], v[100:107], v[64:71]// 00000000249C: CC404040 1D02C9C5
|
|
s_waitcnt lgkmcnt(0) // 0000000024A4: BF89FC07
|
|
s_waitcnt lgkmcnt(0) // 0000000024A8: BF89FC07
|
|
s_barrier // 0000000024AC: BFBD0000
|
|
s_waitcnt lgkmcnt(0) // 0000000024B0: BF89FC07
|
|
v_wmma_f32_16x16x16_f16 v[0:7], v[205:212], v[108:115], v[0:7]// 0000000024B4: CC404000 1C02D9CD
|
|
ds_load_u16 v84, v80 // 0000000024BC: D8F00000 54000050
|
|
ds_load_u16_d16_hi v84, v80 offset:192 // 0000000024C4: DA9C00C0 54000050
|
|
ds_load_u16 v85, v80 offset:384 // 0000000024CC: D8F00180 55000050
|
|
ds_load_u16_d16_hi v85, v80 offset:576 // 0000000024D4: DA9C0240 55000050
|
|
ds_load_u16 v86, v80 offset:768 // 0000000024DC: D8F00300 56000050
|
|
ds_load_u16_d16_hi v86, v80 offset:960 // 0000000024E4: DA9C03C0 56000050
|
|
ds_load_u16 v87, v80 offset:1152 // 0000000024EC: D8F00480 57000050
|
|
v_wmma_f32_16x16x16_f16 v[8:15], v[205:212], v[116:123], v[8:15]// 0000000024F4: CC404008 1C22E9CD
|
|
ds_load_u16_d16_hi v87, v80 offset:1344 // 0000000024FC: DA9C0540 57000050
|
|
ds_load_u16 v88, v80 offset:1536 // 000000002504: D8F00600 58000050
|
|
ds_load_u16_d16_hi v88, v80 offset:1728 // 00000000250C: DA9C06C0 58000050
|
|
ds_load_u16 v89, v80 offset:1920 // 000000002514: D8F00780 59000050
|
|
ds_load_u16_d16_hi v89, v80 offset:2112 // 00000000251C: DA9C0840 59000050
|
|
ds_load_u16 v90, v80 offset:2304 // 000000002524: D8F00900 5A000050
|
|
ds_load_u16_d16_hi v90, v80 offset:2496 // 00000000252C: DA9C09C0 5A000050
|
|
v_wmma_f32_16x16x16_f16 v[16:23], v[205:212], v[124:131], v[16:23]// 000000002534: CC404010 1C42F9CD
|
|
ds_load_u16 v91, v80 offset:2688 // 00000000253C: D8F00A80 5B000050
|
|
ds_load_u16_d16_hi v91, v80 offset:2880 // 000000002544: DA9C0B40 5B000050
|
|
ds_load_b128 v[181:184], v81 // 00000000254C: DBFC0000 B5000051
|
|
ds_load_b128 v[185:188], v81 offset:16 // 000000002554: DBFC0010 B9000051
|
|
ds_load_u16 v92, v80 offset:64 // 00000000255C: D8F00040 5C000050
|
|
ds_load_u16_d16_hi v92, v80 offset:256 // 000000002564: DA9C0100 5C000050
|
|
ds_load_u16 v93, v80 offset:448 // 00000000256C: D8F001C0 5D000050
|
|
v_wmma_f32_16x16x16_f16 v[24:31], v[213:220], v[108:115], v[24:31]// 000000002574: CC404018 1C62D9D5
|
|
ds_load_u16_d16_hi v93, v80 offset:640 // 00000000257C: DA9C0280 5D000050
|
|
ds_load_u16 v94, v80 offset:832 // 000000002584: D8F00340 5E000050
|
|
ds_load_u16_d16_hi v94, v80 offset:1024 // 00000000258C: DA9C0400 5E000050
|
|
ds_load_u16 v95, v80 offset:1216 // 000000002594: D8F004C0 5F000050
|
|
ds_load_u16_d16_hi v95, v80 offset:1408 // 00000000259C: DA9C0580 5F000050
|
|
ds_load_u16 v96, v80 offset:1600 // 0000000025A4: D8F00640 60000050
|
|
ds_load_u16_d16_hi v96, v80 offset:1792 // 0000000025AC: DA9C0700 60000050
|
|
v_wmma_f32_16x16x16_f16 v[32:39], v[213:220], v[116:123], v[32:39]// 0000000025B4: CC404020 1C82E9D5
|
|
ds_load_u16 v97, v80 offset:1984 // 0000000025BC: D8F007C0 61000050
|
|
ds_load_u16_d16_hi v97, v80 offset:2176 // 0000000025C4: DA9C0880 61000050
|
|
ds_load_u16 v98, v80 offset:2368 // 0000000025CC: D8F00940 62000050
|
|
ds_load_u16_d16_hi v98, v80 offset:2560 // 0000000025D4: DA9C0A00 62000050
|
|
ds_load_u16 v99, v80 offset:2752 // 0000000025DC: D8F00AC0 63000050
|
|
ds_load_u16_d16_hi v99, v80 offset:2944 // 0000000025E4: DA9C0B80 63000050
|
|
ds_load_u16 v100, v80 offset:128 // 0000000025EC: D8F00080 64000050
|
|
v_wmma_f32_16x16x16_f16 v[40:47], v[213:220], v[124:131], v[40:47]// 0000000025F4: CC404028 1CA2F9D5
|
|
ds_load_u16_d16_hi v100, v80 offset:320 // 0000000025FC: DA9C0140 64000050
|
|
ds_load_u16 v101, v80 offset:512 // 000000002604: D8F00200 65000050
|
|
ds_load_u16_d16_hi v101, v80 offset:704 // 00000000260C: DA9C02C0 65000050
|
|
ds_load_u16 v102, v80 offset:896 // 000000002614: D8F00380 66000050
|
|
ds_load_u16_d16_hi v102, v80 offset:1088 // 00000000261C: DA9C0440 66000050
|
|
ds_load_u16 v103, v80 offset:1280 // 000000002624: D8F00500 67000050
|
|
ds_load_u16_d16_hi v103, v80 offset:1472 // 00000000262C: DA9C05C0 67000050
|
|
v_wmma_f32_16x16x16_f16 v[48:55], v[221:228], v[108:115], v[48:55]// 000000002634: CC404030 1CC2D9DD
|
|
ds_load_u16 v104, v80 offset:1664 // 00000000263C: D8F00680 68000050
|
|
ds_load_u16_d16_hi v104, v80 offset:1856 // 000000002644: DA9C0740 68000050
|
|
ds_load_u16 v105, v80 offset:2048 // 00000000264C: D8F00800 69000050
|
|
ds_load_u16_d16_hi v105, v80 offset:2240 // 000000002654: DA9C08C0 69000050
|
|
ds_load_u16 v106, v80 offset:2432 // 00000000265C: D8F00980 6A000050
|
|
ds_load_u16_d16_hi v106, v80 offset:2624 // 000000002664: DA9C0A40 6A000050
|
|
v_wmma_f32_16x16x16_f16 v[56:63], v[221:228], v[116:123], v[56:63]// 00000000266C: CC404038 1CE2E9DD
|
|
ds_load_u16 v107, v80 offset:2816 // 000000002674: D8F00B00 6B000050
|
|
ds_load_u16_d16_hi v107, v80 offset:3008 // 00000000267C: DA9C0BC0 6B000050
|
|
ds_load_b128 v[189:192], v81 offset:2560 // 000000002684: DBFC0A00 BD000051
|
|
ds_load_b128 v[193:196], v81 offset:2576 // 00000000268C: DBFC0A10 C1000051
|
|
ds_load_b128 v[197:200], v81 offset:5120 // 000000002694: DBFC1400 C5000051
|
|
ds_load_b128 v[201:204], v81 offset:5136 // 00000000269C: DBFC1410 C9000051
|
|
v_wmma_f32_16x16x16_f16 v[64:71], v[221:228], v[124:131], v[64:71]// 0000000026A4: CC404040 1D02F9DD
|
|
s_sub_u32 s12, s12, 1 // 0000000026AC: 808C810C
|
|
s_cmp_eq_i32 s12, 2 // 0000000026B0: BF00820C
|
|
s_cbranch_scc0 main_loop // 0000000026B4: BFA1FEC3
|
|
s_waitcnt lgkmcnt(4) // 0000000026B8: BF89FC47
|
|
v_wmma_f32_16x16x16_f16 v[0:7], v[181:188], v[84:91], v[0:7]// 0000000026BC: CC404000 1C02A9B5
|
|
ds_load_u16 v108, v80 offset:3104 // 0000000026C4: D8F00C20 6C000050
|
|
ds_load_u16_d16_hi v108, v80 offset:3296 // 0000000026CC: DA9C0CE0 6C000050
|
|
ds_load_u16 v109, v80 offset:3488 // 0000000026D4: D8F00DA0 6D000050
|
|
ds_load_u16_d16_hi v109, v80 offset:3680 // 0000000026DC: DA9C0E60 6D000050
|
|
ds_load_u16 v110, v80 offset:3872 // 0000000026E4: D8F00F20 6E000050
|
|
ds_load_u16_d16_hi v110, v80 offset:4064 // 0000000026EC: DA9C0FE0 6E000050
|
|
ds_load_u16 v111, v80 offset:4256 // 0000000026F4: D8F010A0 6F000050
|
|
s_cmp_eq_u32 s12, s47 // 0000000026FC: BF062F0C
|
|
s_cselect_b32 s66, s60, s64 // 000000002700: 9842403C
|
|
s_cselect_b32 s67, s61, 0 // 000000002704: 9843803D
|
|
v_wmma_f32_16x16x16_f16 v[8:15], v[181:188], v[92:99], v[8:15]// 000000002708: CC404008 1C22B9B5
|
|
ds_load_u16_d16_hi v111, v80 offset:4448 // 000000002710: DA9C1160 6F000050
|
|
ds_load_u16 v112, v80 offset:4640 // 000000002718: D8F01220 70000050
|
|
ds_load_u16_d16_hi v112, v80 offset:4832 // 000000002720: DA9C12E0 70000050
|
|
ds_load_u16 v113, v80 offset:5024 // 000000002728: D8F013A0 71000050
|
|
ds_load_u16_d16_hi v113, v80 offset:5216 // 000000002730: DA9C1460 71000050
|
|
ds_load_u16 v114, v80 offset:5408 // 000000002738: D8F01520 72000050
|
|
ds_load_u16_d16_hi v114, v80 offset:5600 // 000000002740: DA9C15E0 72000050
|
|
s_add_u32 s48, s48, s66 // 000000002748: 80304230
|
|
s_addc_u32 s49, s49, s67 // 00000000274C: 82314331
|
|
s_sub_u32 s56, s56, s66 // 000000002750: 80B84238
|
|
v_wmma_f32_16x16x16_f16 v[16:23], v[181:188], v[100:107], v[16:23]// 000000002754: CC404010 1C42C9B5
|
|
ds_load_u16 v115, v80 offset:5792 // 00000000275C: D8F016A0 73000050
|
|
ds_load_u16_d16_hi v115, v80 offset:5984 // 000000002764: DA9C1760 73000050
|
|
ds_load_b128 v[205:208], v81 offset:32 // 00000000276C: DBFC0020 CD000051
|
|
ds_load_b128 v[209:212], v81 offset:48 // 000000002774: DBFC0030 D1000051
|
|
ds_load_u16 v116, v80 offset:3168 // 00000000277C: D8F00C60 74000050
|
|
ds_load_u16_d16_hi v116, v80 offset:3360 // 000000002784: DA9C0D20 74000050
|
|
ds_load_u16 v117, v80 offset:3552 // 00000000278C: D8F00DE0 75000050
|
|
s_subb_u32 s57, s57, s67 // 000000002794: 82B94339
|
|
s_cmp_eq_u32 s57, 0 // 000000002798: BF068039
|
|
s_cselect_b32 s50, s56, -1 // 00000000279C: 9832C138
|
|
s_waitcnt lgkmcnt(21) // 0000000027A0: BF89FD57
|
|
v_wmma_f32_16x16x16_f16 v[24:31], v[189:196], v[84:91], v[24:31]// 0000000027A4: CC404018 1C62A9BD
|
|
ds_load_u16_d16_hi v117, v80 offset:3744 // 0000000027AC: DA9C0EA0 75000050
|
|
ds_load_u16 v118, v80 offset:3936 // 0000000027B4: D8F00F60 76000050
|
|
ds_load_u16_d16_hi v118, v80 offset:4128 // 0000000027BC: DA9C1020 76000050
|
|
ds_load_u16 v119, v80 offset:4320 // 0000000027C4: D8F010E0 77000050
|
|
ds_load_u16_d16_hi v119, v80 offset:4512 // 0000000027CC: DA9C11A0 77000050
|
|
ds_load_u16 v120, v80 offset:4704 // 0000000027D4: D8F01260 78000050
|
|
ds_load_u16_d16_hi v120, v80 offset:4896 // 0000000027DC: DA9C1320 78000050
|
|
s_cmp_eq_u32 s12, s47 // 0000000027E4: BF062F0C
|
|
s_cselect_b32 s66, s62, s65 // 0000000027E8: 9842413E
|
|
s_cselect_b32 s67, s63, 0 // 0000000027EC: 9843803F
|
|
v_wmma_f32_16x16x16_f16 v[32:39], v[189:196], v[92:99], v[32:39]// 0000000027F0: CC404020 1C82B9BD
|
|
ds_load_u16 v121, v80 offset:5088 // 0000000027F8: D8F013E0 79000050
|
|
ds_load_u16_d16_hi v121, v80 offset:5280 // 000000002800: DA9C14A0 79000050
|
|
ds_load_u16 v122, v80 offset:5472 // 000000002808: D8F01560 7A000050
|
|
ds_load_u16_d16_hi v122, v80 offset:5664 // 000000002810: DA9C1620 7A000050
|
|
ds_load_u16 v123, v80 offset:5856 // 000000002818: D8F016E0 7B000050
|
|
ds_load_u16_d16_hi v123, v80 offset:6048 // 000000002820: DA9C17A0 7B000050
|
|
ds_load_u16 v124, v80 offset:3232 // 000000002828: D8F00CA0 7C000050
|
|
s_add_u32 s52, s52, s66 // 000000002830: 80344234
|
|
s_addc_u32 s53, s53, s67 // 000000002834: 82354335
|
|
s_sub_u32 s58, s58, s66 // 000000002838: 80BA423A
|
|
v_wmma_f32_16x16x16_f16 v[40:47], v[189:196], v[100:107], v[40:47]// 00000000283C: CC404028 1CA2C9BD
|
|
ds_load_u16_d16_hi v124, v80 offset:3424 // 000000002844: DA9C0D60 7C000050
|
|
ds_load_u16 v125, v80 offset:3616 // 00000000284C: D8F00E20 7D000050
|
|
ds_load_u16_d16_hi v125, v80 offset:3808 // 000000002854: DA9C0EE0 7D000050
|
|
ds_load_u16 v126, v80 offset:4000 // 00000000285C: D8F00FA0 7E000050
|
|
ds_load_u16_d16_hi v126, v80 offset:4192 // 000000002864: DA9C1060 7E000050
|
|
ds_load_u16 v127, v80 offset:4384 // 00000000286C: D8F01120 7F000050
|
|
ds_load_u16_d16_hi v127, v80 offset:4576 // 000000002874: DA9C11E0 7F000050
|
|
s_subb_u32 s59, s59, s67 // 00000000287C: 82BB433B
|
|
s_cmp_eq_u32 s59, 0 // 000000002880: BF06803B
|
|
s_cselect_b32 s54, s58, -1 // 000000002884: 9836C13A
|
|
s_waitcnt vmcnt(5) // 000000002888: BF8917F7
|
|
ds_store_b128 v78, v[230:233] // 00000000288C: DB7C0000 0000E64E
|
|
s_waitcnt vmcnt(4) // 000000002894: BF8913F7
|
|
ds_store_b128 v78, v[234:237] offset:64 // 000000002898: DB7C0040 0000EA4E
|
|
s_waitcnt vmcnt(3) // 0000000028A0: BF890FF7
|
|
ds_store_b128 v78, v[238:241] offset:128 // 0000000028A4: DB7C0080 0000EE4E
|
|
s_waitcnt vmcnt(2) // 0000000028AC: BF890BF7
|
|
ds_store_b128 v79, v[242:245] // 0000000028B0: DB7C0000 0000F24F
|
|
s_waitcnt vmcnt(1) // 0000000028B8: BF8907F7
|
|
ds_store_b128 v79, v[246:249] offset:2560 // 0000000028BC: DB7C0A00 0000F64F
|
|
s_waitcnt vmcnt(0) // 0000000028C4: BF8903F7
|
|
ds_store_b128 v79, v[250:253] offset:5120 // 0000000028C8: DB7C1400 0000FA4F
|
|
v_xor_b32_e32 v78, 0x4000, v78 // 0000000028D0: 3A9C9CFF 00004000
|
|
v_xor_b32_e32 v79, 0x4000, v79 // 0000000028D8: 3A9E9EFF 00004000
|
|
v_wmma_f32_16x16x16_f16 v[48:55], v[197:204], v[84:91], v[48:55]// 0000000028E0: CC404030 1CC2A9C5
|
|
ds_load_u16 v128, v80 offset:4768 // 0000000028E8: D8F012A0 80000050
|
|
ds_load_u16_d16_hi v128, v80 offset:4960 // 0000000028F0: DA9C1360 80000050
|
|
ds_load_u16 v129, v80 offset:5152 // 0000000028F8: D8F01420 81000050
|
|
ds_load_u16_d16_hi v129, v80 offset:5344 // 000000002900: DA9C14E0 81000050
|
|
ds_load_u16 v130, v80 offset:5536 // 000000002908: D8F015A0 82000050
|
|
ds_load_u16_d16_hi v130, v80 offset:5728 // 000000002910: DA9C1660 82000050
|
|
v_wmma_f32_16x16x16_f16 v[56:63], v[197:204], v[92:99], v[56:63]// 000000002918: CC404038 1CE2B9C5
|
|
ds_load_u16 v131, v80 offset:5920 // 000000002920: D8F01720 83000050
|
|
ds_load_u16_d16_hi v131, v80 offset:6112 // 000000002928: DA9C17E0 83000050
|
|
ds_load_b128 v[213:216], v81 offset:2592 // 000000002930: DBFC0A20 D5000051
|
|
ds_load_b128 v[217:220], v81 offset:2608 // 000000002938: DBFC0A30 D9000051
|
|
ds_load_b128 v[221:224], v81 offset:5152 // 000000002940: DBFC1420 DD000051
|
|
ds_load_b128 v[225:228], v81 offset:5168 // 000000002948: DBFC1430 E1000051
|
|
v_xor_b32_e32 v80, 0x4000, v80 // 000000002950: 3AA0A0FF 00004000
|
|
v_xor_b32_e32 v81, 0x4000, v81 // 000000002958: 3AA2A2FF 00004000
|
|
v_wmma_f32_16x16x16_f16 v[64:71], v[197:204], v[100:107], v[64:71]// 000000002960: CC404040 1D02C9C5
|
|
s_waitcnt lgkmcnt(0) // 000000002968: BF89FC07
|
|
s_waitcnt lgkmcnt(0) // 00000000296C: BF89FC07
|
|
s_barrier // 000000002970: BFBD0000
|
|
s_waitcnt lgkmcnt(0) // 000000002974: BF89FC07
|
|
v_wmma_f32_16x16x16_f16 v[0:7], v[205:212], v[108:115], v[0:7]// 000000002978: CC404000 1C02D9CD
|
|
ds_load_u16 v84, v80 // 000000002980: D8F00000 54000050
|
|
ds_load_u16_d16_hi v84, v80 offset:192 // 000000002988: DA9C00C0 54000050
|
|
ds_load_u16 v85, v80 offset:384 // 000000002990: D8F00180 55000050
|
|
ds_load_u16_d16_hi v85, v80 offset:576 // 000000002998: DA9C0240 55000050
|
|
ds_load_u16 v86, v80 offset:768 // 0000000029A0: D8F00300 56000050
|
|
ds_load_u16_d16_hi v86, v80 offset:960 // 0000000029A8: DA9C03C0 56000050
|
|
ds_load_u16 v87, v80 offset:1152 // 0000000029B0: D8F00480 57000050
|
|
v_wmma_f32_16x16x16_f16 v[8:15], v[205:212], v[116:123], v[8:15]// 0000000029B8: CC404008 1C22E9CD
|
|
ds_load_u16_d16_hi v87, v80 offset:1344 // 0000000029C0: DA9C0540 57000050
|
|
ds_load_u16 v88, v80 offset:1536 // 0000000029C8: D8F00600 58000050
|
|
ds_load_u16_d16_hi v88, v80 offset:1728 // 0000000029D0: DA9C06C0 58000050
|
|
ds_load_u16 v89, v80 offset:1920 // 0000000029D8: D8F00780 59000050
|
|
ds_load_u16_d16_hi v89, v80 offset:2112 // 0000000029E0: DA9C0840 59000050
|
|
ds_load_u16 v90, v80 offset:2304 // 0000000029E8: D8F00900 5A000050
|
|
ds_load_u16_d16_hi v90, v80 offset:2496 // 0000000029F0: DA9C09C0 5A000050
|
|
v_wmma_f32_16x16x16_f16 v[16:23], v[205:212], v[124:131], v[16:23]// 0000000029F8: CC404010 1C42F9CD
|
|
ds_load_u16 v91, v80 offset:2688 // 000000002A00: D8F00A80 5B000050
|
|
ds_load_u16_d16_hi v91, v80 offset:2880 // 000000002A08: DA9C0B40 5B000050
|
|
ds_load_b128 v[181:184], v81 // 000000002A10: DBFC0000 B5000051
|
|
ds_load_b128 v[185:188], v81 offset:16 // 000000002A18: DBFC0010 B9000051
|
|
ds_load_u16 v92, v80 offset:64 // 000000002A20: D8F00040 5C000050
|
|
ds_load_u16_d16_hi v92, v80 offset:256 // 000000002A28: DA9C0100 5C000050
|
|
ds_load_u16 v93, v80 offset:448 // 000000002A30: D8F001C0 5D000050
|
|
v_wmma_f32_16x16x16_f16 v[24:31], v[213:220], v[108:115], v[24:31]// 000000002A38: CC404018 1C62D9D5
|
|
ds_load_u16_d16_hi v93, v80 offset:640 // 000000002A40: DA9C0280 5D000050
|
|
ds_load_u16 v94, v80 offset:832 // 000000002A48: D8F00340 5E000050
|
|
ds_load_u16_d16_hi v94, v80 offset:1024 // 000000002A50: DA9C0400 5E000050
|
|
ds_load_u16 v95, v80 offset:1216 // 000000002A58: D8F004C0 5F000050
|
|
ds_load_u16_d16_hi v95, v80 offset:1408 // 000000002A60: DA9C0580 5F000050
|
|
ds_load_u16 v96, v80 offset:1600 // 000000002A68: D8F00640 60000050
|
|
ds_load_u16_d16_hi v96, v80 offset:1792 // 000000002A70: DA9C0700 60000050
|
|
v_wmma_f32_16x16x16_f16 v[32:39], v[213:220], v[116:123], v[32:39]// 000000002A78: CC404020 1C82E9D5
|
|
ds_load_u16 v97, v80 offset:1984 // 000000002A80: D8F007C0 61000050
|
|
ds_load_u16_d16_hi v97, v80 offset:2176 // 000000002A88: DA9C0880 61000050
|
|
ds_load_u16 v98, v80 offset:2368 // 000000002A90: D8F00940 62000050
|
|
ds_load_u16_d16_hi v98, v80 offset:2560 // 000000002A98: DA9C0A00 62000050
|
|
ds_load_u16 v99, v80 offset:2752 // 000000002AA0: D8F00AC0 63000050
|
|
ds_load_u16_d16_hi v99, v80 offset:2944 // 000000002AA8: DA9C0B80 63000050
|
|
ds_load_u16 v100, v80 offset:128 // 000000002AB0: D8F00080 64000050
|
|
v_wmma_f32_16x16x16_f16 v[40:47], v[213:220], v[124:131], v[40:47]// 000000002AB8: CC404028 1CA2F9D5
|
|
ds_load_u16_d16_hi v100, v80 offset:320 // 000000002AC0: DA9C0140 64000050
|
|
ds_load_u16 v101, v80 offset:512 // 000000002AC8: D8F00200 65000050
|
|
ds_load_u16_d16_hi v101, v80 offset:704 // 000000002AD0: DA9C02C0 65000050
|
|
ds_load_u16 v102, v80 offset:896 // 000000002AD8: D8F00380 66000050
|
|
ds_load_u16_d16_hi v102, v80 offset:1088 // 000000002AE0: DA9C0440 66000050
|
|
ds_load_u16 v103, v80 offset:1280 // 000000002AE8: D8F00500 67000050
|
|
ds_load_u16_d16_hi v103, v80 offset:1472 // 000000002AF0: DA9C05C0 67000050
|
|
v_wmma_f32_16x16x16_f16 v[48:55], v[221:228], v[108:115], v[48:55]// 000000002AF8: CC404030 1CC2D9DD
|
|
ds_load_u16 v104, v80 offset:1664 // 000000002B00: D8F00680 68000050
|
|
ds_load_u16_d16_hi v104, v80 offset:1856 // 000000002B08: DA9C0740 68000050
|
|
ds_load_u16 v105, v80 offset:2048 // 000000002B10: D8F00800 69000050
|
|
ds_load_u16_d16_hi v105, v80 offset:2240 // 000000002B18: DA9C08C0 69000050
|
|
ds_load_u16 v106, v80 offset:2432 // 000000002B20: D8F00980 6A000050
|
|
ds_load_u16_d16_hi v106, v80 offset:2624 // 000000002B28: DA9C0A40 6A000050
|
|
v_wmma_f32_16x16x16_f16 v[56:63], v[221:228], v[116:123], v[56:63]// 000000002B30: CC404038 1CE2E9DD
|
|
ds_load_u16 v107, v80 offset:2816 // 000000002B38: D8F00B00 6B000050
|
|
ds_load_u16_d16_hi v107, v80 offset:3008 // 000000002B40: DA9C0BC0 6B000050
|
|
ds_load_b128 v[189:192], v81 offset:2560 // 000000002B48: DBFC0A00 BD000051
|
|
ds_load_b128 v[193:196], v81 offset:2576 // 000000002B50: DBFC0A10 C1000051
|
|
ds_load_b128 v[197:200], v81 offset:5120 // 000000002B58: DBFC1400 C5000051
|
|
ds_load_b128 v[201:204], v81 offset:5136 // 000000002B60: DBFC1410 C9000051
|
|
v_wmma_f32_16x16x16_f16 v[64:71], v[221:228], v[124:131], v[64:71]// 000000002B68: CC404040 1D02F9DD
|
|
s_and_b32 s8, s46, 0x3fff // 000000002B70: 8B08FF2E 00003FFF
|
|
s_mov_b32 s69, 0 // 000000002B78: BEC50080
|
|
s_mul_i32 s68, 0x555, s24 // 000000002B7C: 964418FF 00000555
|
|
s_lshl_b64 s[68:69], s[68:69], 16 // 000000002B84: 84C49044
|
|
s_mul_i32 s67, s24, 0x5556 // 000000002B88: 9643FF18 00005556
|
|
s_add_u32 s68, s67, s68 // 000000002B90: 80444443
|
|
s_addc_u32 s69, s69, 0 // 000000002B94: 82458045
|
|
s_lshr_b64 s[68:69], s[68:69], 33 // 000000002B98: 85C4A144
|
|
s_mov_b32 s67, s68 // 000000002B9C: BEC30044
|
|
s_mul_i32 s68, s67, 0x60 // 000000002BA0: 9644FF43 00000060
|
|
s_sub_u32 s66, s24, s68 // 000000002BA8: 80C24418
|
|
s_add_u32 s67, -1, s14 // 000000002BAC: 80430EC1
|
|
s_cmp_ge_u32 s2, s67 // 000000002BB0: BF094302
|
|
s_cselect_b32 s66, s66, 0 // 000000002BB4: 98428042
|
|
s_cmpk_gt_u32 s66, 0x0 // 000000002BB8: B5C20000
|
|
s_cbranch_scc1 epilogue // 000000002BBC: BFA2035F
|
|
s_mov_b32 s69, 0 // 000000002BC0: BEC50080
|
|
s_mul_i32 s68, 0x555, s25 // 000000002BC4: 964419FF 00000555
|
|
s_lshl_b64 s[68:69], s[68:69], 16 // 000000002BCC: 84C49044
|
|
s_mul_i32 s67, s25, 0x5556 // 000000002BD0: 9643FF19 00005556
|
|
s_add_u32 s68, s67, s68 // 000000002BD8: 80444443
|
|
s_addc_u32 s69, s69, 0 // 000000002BDC: 82458045
|
|
s_lshr_b64 s[68:69], s[68:69], 33 // 000000002BE0: 85C4A144
|
|
s_mov_b32 s67, s68 // 000000002BE4: BEC30044
|
|
s_mul_i32 s68, s67, 0x60 // 000000002BE8: 9644FF43 00000060
|
|
s_sub_u32 s66, s25, s68 // 000000002BF0: 80C24419
|
|
s_add_u32 s67, -1, s15 // 000000002BF4: 80430FC1
|
|
s_cmp_ge_u32 s3, s67 // 000000002BF8: BF094303
|
|
s_cselect_b32 s66, s66, 0 // 000000002BFC: 98428042
|
|
s_cmpk_gt_u32 s66, 0x0 // 000000002C00: B5C20000
|
|
s_cbranch_scc1 epilogue // 000000002C04: BFA2034D
|
|
s_and_b32 s67, 31, s27 // 000000002C08: 8B431B9F
|
|
s_waitcnt lgkmcnt(4) // 000000002C0C: BF89FC47
|
|
v_wmma_f32_16x16x16_f16 v[0:7], v[181:188], v[84:91], v[0:7]// 000000002C10: CC404000 1C02A9B5
|
|
ds_load_u16 v108, v80 offset:3104 // 000000002C18: D8F00C20 6C000050
|
|
ds_load_u16_d16_hi v108, v80 offset:3296 // 000000002C20: DA9C0CE0 6C000050
|
|
ds_load_u16 v109, v80 offset:3488 // 000000002C28: D8F00DA0 6D000050
|
|
ds_load_u16_d16_hi v109, v80 offset:3680 // 000000002C30: DA9C0E60 6D000050
|
|
ds_load_u16 v110, v80 offset:3872 // 000000002C38: D8F00F20 6E000050
|
|
ds_load_u16_d16_hi v110, v80 offset:4064 // 000000002C40: DA9C0FE0 6E000050
|
|
ds_load_u16 v111, v80 offset:4256 // 000000002C48: D8F010A0 6F000050
|
|
v_wmma_f32_16x16x16_f16 v[8:15], v[181:188], v[92:99], v[8:15]// 000000002C50: CC404008 1C22B9B5
|
|
ds_load_u16_d16_hi v111, v80 offset:4448 // 000000002C58: DA9C1160 6F000050
|
|
ds_load_u16 v112, v80 offset:4640 // 000000002C60: D8F01220 70000050
|
|
ds_load_u16_d16_hi v112, v80 offset:4832 // 000000002C68: DA9C12E0 70000050
|
|
ds_load_u16 v113, v80 offset:5024 // 000000002C70: D8F013A0 71000050
|
|
ds_load_u16_d16_hi v113, v80 offset:5216 // 000000002C78: DA9C1460 71000050
|
|
ds_load_u16 v114, v80 offset:5408 // 000000002C80: D8F01520 72000050
|
|
ds_load_u16_d16_hi v114, v80 offset:5600 // 000000002C88: DA9C15E0 72000050
|
|
v_wmma_f32_16x16x16_f16 v[16:23], v[181:188], v[100:107], v[16:23]// 000000002C90: CC404010 1C42C9B5
|
|
ds_load_u16 v115, v80 offset:5792 // 000000002C98: D8F016A0 73000050
|
|
ds_load_u16_d16_hi v115, v80 offset:5984 // 000000002CA0: DA9C1760 73000050
|
|
ds_load_b128 v[205:208], v81 offset:32 // 000000002CA8: DBFC0020 CD000051
|
|
ds_load_b128 v[209:212], v81 offset:48 // 000000002CB0: DBFC0030 D1000051
|
|
ds_load_u16 v116, v80 offset:3168 // 000000002CB8: D8F00C60 74000050
|
|
ds_load_u16_d16_hi v116, v80 offset:3360 // 000000002CC0: DA9C0D20 74000050
|
|
ds_load_u16 v117, v80 offset:3552 // 000000002CC8: D8F00DE0 75000050
|
|
s_waitcnt lgkmcnt(21) // 000000002CD0: BF89FD57
|
|
v_wmma_f32_16x16x16_f16 v[24:31], v[189:196], v[84:91], v[24:31]// 000000002CD4: CC404018 1C62A9BD
|
|
ds_load_u16_d16_hi v117, v80 offset:3744 // 000000002CDC: DA9C0EA0 75000050
|
|
ds_load_u16 v118, v80 offset:3936 // 000000002CE4: D8F00F60 76000050
|
|
ds_load_u16_d16_hi v118, v80 offset:4128 // 000000002CEC: DA9C1020 76000050
|
|
ds_load_u16 v119, v80 offset:4320 // 000000002CF4: D8F010E0 77000050
|
|
ds_load_u16_d16_hi v119, v80 offset:4512 // 000000002CFC: DA9C11A0 77000050
|
|
ds_load_u16 v120, v80 offset:4704 // 000000002D04: D8F01260 78000050
|
|
ds_load_u16_d16_hi v120, v80 offset:4896 // 000000002D0C: DA9C1320 78000050
|
|
v_wmma_f32_16x16x16_f16 v[32:39], v[189:196], v[92:99], v[32:39]// 000000002D14: CC404020 1C82B9BD
|
|
ds_load_u16 v121, v80 offset:5088 // 000000002D1C: D8F013E0 79000050
|
|
ds_load_u16_d16_hi v121, v80 offset:5280 // 000000002D24: DA9C14A0 79000050
|
|
ds_load_u16 v122, v80 offset:5472 // 000000002D2C: D8F01560 7A000050
|
|
ds_load_u16_d16_hi v122, v80 offset:5664 // 000000002D34: DA9C1620 7A000050
|
|
ds_load_u16 v123, v80 offset:5856 // 000000002D3C: D8F016E0 7B000050
|
|
ds_load_u16_d16_hi v123, v80 offset:6048 // 000000002D44: DA9C17A0 7B000050
|
|
ds_load_u16 v124, v80 offset:3232 // 000000002D4C: D8F00CA0 7C000050
|
|
v_wmma_f32_16x16x16_f16 v[40:47], v[189:196], v[100:107], v[40:47]// 000000002D54: CC404028 1CA2C9BD
|
|
ds_load_u16_d16_hi v124, v80 offset:3424 // 000000002D5C: DA9C0D60 7C000050
|
|
ds_load_u16 v125, v80 offset:3616 // 000000002D64: D8F00E20 7D000050
|
|
ds_load_u16_d16_hi v125, v80 offset:3808 // 000000002D6C: DA9C0EE0 7D000050
|
|
ds_load_u16 v126, v80 offset:4000 // 000000002D74: D8F00FA0 7E000050
|
|
ds_load_u16_d16_hi v126, v80 offset:4192 // 000000002D7C: DA9C1060 7E000050
|
|
ds_load_u16 v127, v80 offset:4384 // 000000002D84: D8F01120 7F000050
|
|
ds_load_u16_d16_hi v127, v80 offset:4576 // 000000002D8C: DA9C11E0 7F000050
|
|
v_wmma_f32_16x16x16_f16 v[48:55], v[197:204], v[84:91], v[48:55]// 000000002D94: CC404030 1CC2A9C5
|
|
ds_load_u16 v128, v80 offset:4768 // 000000002D9C: D8F012A0 80000050
|
|
ds_load_u16_d16_hi v128, v80 offset:4960 // 000000002DA4: DA9C1360 80000050
|
|
ds_load_u16 v129, v80 offset:5152 // 000000002DAC: D8F01420 81000050
|
|
ds_load_u16_d16_hi v129, v80 offset:5344 // 000000002DB4: DA9C14E0 81000050
|
|
ds_load_u16 v130, v80 offset:5536 // 000000002DBC: D8F015A0 82000050
|
|
ds_load_u16_d16_hi v130, v80 offset:5728 // 000000002DC4: DA9C1660 82000050
|
|
v_wmma_f32_16x16x16_f16 v[56:63], v[197:204], v[92:99], v[56:63]// 000000002DCC: CC404038 1CE2B9C5
|
|
ds_load_u16 v131, v80 offset:5920 // 000000002DD4: D8F01720 83000050
|
|
ds_load_u16_d16_hi v131, v80 offset:6112 // 000000002DDC: DA9C17E0 83000050
|
|
ds_load_b128 v[213:216], v81 offset:2592 // 000000002DE4: DBFC0A20 D5000051
|
|
ds_load_b128 v[217:220], v81 offset:2608 // 000000002DEC: DBFC0A30 D9000051
|
|
ds_load_b128 v[221:224], v81 offset:5152 // 000000002DF4: DBFC1420 DD000051
|
|
ds_load_b128 v[225:228], v81 offset:5168 // 000000002DFC: DBFC1430 E1000051
|
|
v_wmma_f32_16x16x16_f16 v[64:71], v[197:204], v[100:107], v[64:71]// 000000002E04: CC404040 1D02C9C5
|
|
s_waitcnt lgkmcnt(0) // 000000002E0C: BF89FC07
|
|
v_wmma_f32_16x16x16_f16 v[0:7], v[205:212], v[108:115], v[0:7]// 000000002E10: CC404000 1C02D9CD
|
|
v_wmma_f32_16x16x16_f16 v[8:15], v[205:212], v[116:123], v[8:15]// 000000002E18: CC404008 1C22E9CD
|
|
v_wmma_f32_16x16x16_f16 v[16:23], v[205:212], v[124:131], v[16:23]// 000000002E20: CC404010 1C42F9CD
|
|
v_wmma_f32_16x16x16_f16 v[24:31], v[213:220], v[108:115], v[24:31]// 000000002E28: CC404018 1C62D9D5
|
|
v_wmma_f32_16x16x16_f16 v[32:39], v[213:220], v[116:123], v[32:39]// 000000002E30: CC404020 1C82E9D5
|
|
v_wmma_f32_16x16x16_f16 v[40:47], v[213:220], v[124:131], v[40:47]// 000000002E38: CC404028 1CA2F9D5
|
|
v_wmma_f32_16x16x16_f16 v[48:55], v[221:228], v[108:115], v[48:55]// 000000002E40: CC404030 1CC2D9DD
|
|
v_wmma_f32_16x16x16_f16 v[56:63], v[221:228], v[116:123], v[56:63]// 000000002E48: CC404038 1CE2E9DD
|
|
v_wmma_f32_16x16x16_f16 v[64:71], v[221:228], v[124:131], v[64:71]// 000000002E50: CC404040 1D02F9DD
|
|
s_load_b256 s[48:55], s[0:1], 0x58 // 000000002E58: F40C0C00 F8000058
|
|
s_load_b32 s56, s[0:1], 0x78 // 000000002E60: F4000E00 F8000078
|
|
v_lshrrev_b32_e32 v76, 5, v254 // 000000002E68: 3299FC85
|
|
v_lshrrev_b32_e32 v77, 1, v76 // 000000002E6C: 329A9881
|
|
v_mul_lo_u32 v77, 16, v77 // 000000002E70: D72C004D 00029A90
|
|
v_and_b32_e32 v73, 31, v254 // 000000002E78: 3693FC9F
|
|
v_lshrrev_b32_e32 v73, 4, v73 // 000000002E7C: 32929284
|
|
v_add_lshl_u32 v73, v77, v73, 0 // 000000002E80: D6470049 0202934D
|
|
v_mul_lo_u32 v74, v73, s38 // 000000002E88: D72C004A 00004D49
|
|
v_mul_lo_u32 v75, v73, s36 // 000000002E90: D72C004B 00004949
|
|
v_and_b32_e32 v72, 1, v76 // 000000002E98: 36909881
|
|
v_mul_lo_u32 v72, 16, v72 // 000000002E9C: D72C0048 00029090
|
|
v_and_b32_e32 v77, 15, v254 // 000000002EA4: 369BFC8F
|
|
v_add_lshl_u32 v72, v77, v72, 0 // 000000002EA8: D6470048 0202914D
|
|
s_mul_i32 s8, 0x60, s2 // 000000002EB0: 960802FF 00000060
|
|
v_add_nc_u32_e32 v72, s8, v72 // 000000002EB8: 4A909008
|
|
s_mul_i32 s8, 0x60, s3 // 000000002EBC: 960803FF 00000060
|
|
v_add_nc_u32_e32 v73, s8, v73 // 000000002EC4: 4A929208
|
|
s_waitcnt lgkmcnt(0) // 000000002EC8: BF89FC07
|
|
s_mov_b64 s[32:33], s[48:49] // 000000002ECC: BEA00130
|
|
s_mov_b32 s35, 0x31004000 // 000000002ED0: BEA300FF 31004000
|
|
s_mov_b32 s34, 0 // 000000002ED8: BEA20080
|
|
s_mul_i32 s34, 4, s34 // 000000002EDC: 96222284
|
|
s_add_u32 s8, s4, 1 // 000000002EE0: 80088104
|
|
s_mul_i32 s8, s53, s8 // 000000002EE4: 96080835
|
|
s_cmp_eq_u32 s8, 0 // 000000002EE8: BF068008
|
|
s_cselect_b32 s8, s24, s8 // 000000002EEC: 98080818
|
|
s_mov_b64 s[40:41], s[50:51] // 000000002EF0: BEA80132
|
|
s_mov_b32 s43, 0x31004000 // 000000002EF4: BEAB00FF 31004000
|
|
s_mov_b32 s42, 0 // 000000002EFC: BEAA0080
|
|
s_mul_i32 s8, 0x60, s2 // 000000002F00: 960802FF 00000060
|
|
v_add_nc_u32_e32 v80, s8, v254 // 000000002F08: 4AA1FC08
|
|
s_mul_i32 s42, 4, s42 // 000000002F0C: 962A2A84
|
|
s_mul_i32 s8, s53, s4 // 000000002F10: 96080435
|
|
v_add_nc_u32_e32 v78, s8, v80 // 000000002F14: 4A9CA008
|
|
v_lshlrev_b32_e32 v78, 2, v78 // 000000002F18: 309C9C82
|
|
v_lshlrev_b32_e32 v79, 2, v80 // 000000002F1C: 309EA082
|
|
s_mul_i32 s8, 0x60, s3 // 000000002F20: 960803FF 00000060
|
|
v_add_nc_u32_e32 v80, s8, v254 // 000000002F28: 4AA1FC08
|
|
buffer_load_b32 v76, v78, s[40:43], 0 offen // 000000002F2C: E0500000 804A4C4E
|
|
buffer_load_b32 v77, v79, s[32:35], 0 offen // 000000002F34: E0500000 80484D4F
|
|
v_lshlrev_b32_e32 v80, 2, v254 // 000000002F3C: 30A1FC82
|
|
s_barrier // 000000002F40: BFBD0000
|
|
s_waitcnt vmcnt(1) // 000000002F44: BF8907F7
|
|
ds_store_b32 v80, v76 // 000000002F48: D8340000 00004C50
|
|
v_cmp_gt_u32_e64 s48, s34, 0 // 000000002F50: D44C0030 00010022
|
|
s_waitcnt vmcnt(0) // 000000002F58: BF8903F7
|
|
v_cndmask_b32_e64 v77, 1.0, v77, s48 // 000000002F5C: D501004D 00C29AF2
|
|
ds_store_b32 v80, v77 offset:512 // 000000002F64: D8340200 00004D50
|
|
s_add_u32 s12, s12, 0xafac // 000000002F6C: 800CFF0C 0000AFAC
|
|
s_addc_u32 s13, s13, 0 // 000000002F74: 820D800D
|
|
s_mul_i32 s8, 0x60, s2 // 000000002F78: 960802FF 00000060
|
|
v_sub_nc_u32_e64 v81, v72, s8 // 000000002F80: D5260051 00001148
|
|
v_lshlrev_b32_e32 v81, 2, v81 // 000000002F88: 30A2A282
|
|
s_waitcnt lgkmcnt(0) // 000000002F8C: BF89FC07
|
|
s_barrier // 000000002F90: BFBD0000
|
|
ds_load_b32 v138, v81 // 000000002F94: D8D80000 8A000051
|
|
ds_load_b32 v139, v81 offset:512 // 000000002F9C: D8D80200 8B000051
|
|
ds_load_b32 v140, v81 offset:128 // 000000002FA4: D8D80080 8C000051
|
|
ds_load_b32 v141, v81 offset:640 // 000000002FAC: D8D80280 8D000051
|
|
ds_load_b32 v142, v81 offset:256 // 000000002FB4: D8D80100 8E000051
|
|
ds_load_b32 v143, v81 offset:768 // 000000002FBC: D8D80300 8F000051
|
|
v_add_lshl_u32 v79, v75, v72, 1 // 000000002FC4: D647004F 0206914B
|
|
v_mov_b32_e32 v82, v0 // 000000002FCC: 7EA40300
|
|
v_mov_b32_e32 v83, v8 // 000000002FD0: 7EA60308
|
|
v_mov_b32_e32 v84, v16 // 000000002FD4: 7EA80310
|
|
v_mov_b32_e32 v85, v1 // 000000002FD8: 7EAA0301
|
|
v_mov_b32_e32 v86, v9 // 000000002FDC: 7EAC0309
|
|
v_mov_b32_e32 v87, v17 // 000000002FE0: 7EAE0311
|
|
v_mov_b32_e32 v88, v2 // 000000002FE4: 7EB00302
|
|
v_mov_b32_e32 v89, v10 // 000000002FE8: 7EB2030A
|
|
v_mov_b32_e32 v90, v18 // 000000002FEC: 7EB40312
|
|
v_mov_b32_e32 v91, v3 // 000000002FF0: 7EB60303
|
|
v_mov_b32_e32 v92, v11 // 000000002FF4: 7EB8030B
|
|
v_mov_b32_e32 v93, v19 // 000000002FF8: 7EBA0313
|
|
v_mov_b32_e32 v94, v4 // 000000002FFC: 7EBC0304
|
|
v_mov_b32_e32 v95, v12 // 000000003000: 7EBE030C
|
|
v_mov_b32_e32 v96, v20 // 000000003004: 7EC00314
|
|
v_mov_b32_e32 v97, v5 // 000000003008: 7EC20305
|
|
v_mov_b32_e32 v98, v13 // 00000000300C: 7EC4030D
|
|
v_mov_b32_e32 v99, v21 // 000000003010: 7EC60315
|
|
v_mov_b32_e32 v100, v6 // 000000003014: 7EC80306
|
|
v_mov_b32_e32 v101, v14 // 000000003018: 7ECA030E
|
|
v_mov_b32_e32 v102, v22 // 00000000301C: 7ECC0316
|
|
v_mov_b32_e32 v103, v7 // 000000003020: 7ECE0307
|
|
v_mov_b32_e32 v104, v15 // 000000003024: 7ED0030F
|
|
v_mov_b32_e32 v105, v23 // 000000003028: 7ED20317
|
|
v_mov_b32_e32 v106, v24 // 00000000302C: 7ED40318
|
|
v_mov_b32_e32 v107, v32 // 000000003030: 7ED60320
|
|
v_mov_b32_e32 v108, v40 // 000000003034: 7ED80328
|
|
v_mov_b32_e32 v109, v25 // 000000003038: 7EDA0319
|
|
v_mov_b32_e32 v110, v33 // 00000000303C: 7EDC0321
|
|
v_mov_b32_e32 v111, v41 // 000000003040: 7EDE0329
|
|
v_mov_b32_e32 v112, v26 // 000000003044: 7EE0031A
|
|
v_mov_b32_e32 v113, v34 // 000000003048: 7EE20322
|
|
v_mov_b32_e32 v114, v42 // 00000000304C: 7EE4032A
|
|
v_mov_b32_e32 v115, v27 // 000000003050: 7EE6031B
|
|
v_mov_b32_e32 v116, v35 // 000000003054: 7EE80323
|
|
v_mov_b32_e32 v117, v43 // 000000003058: 7EEA032B
|
|
v_mov_b32_e32 v118, v28 // 00000000305C: 7EEC031C
|
|
v_mov_b32_e32 v119, v36 // 000000003060: 7EEE0324
|
|
v_mov_b32_e32 v120, v44 // 000000003064: 7EF0032C
|
|
v_mov_b32_e32 v121, v29 // 000000003068: 7EF2031D
|
|
v_mov_b32_e32 v122, v37 // 00000000306C: 7EF40325
|
|
v_mov_b32_e32 v123, v45 // 000000003070: 7EF6032D
|
|
v_mov_b32_e32 v124, v30 // 000000003074: 7EF8031E
|
|
v_mov_b32_e32 v125, v38 // 000000003078: 7EFA0326
|
|
v_mov_b32_e32 v126, v46 // 00000000307C: 7EFC032E
|
|
v_mov_b32_e32 v127, v31 // 000000003080: 7EFE031F
|
|
v_mov_b32_e32 v128, v39 // 000000003084: 7F000327
|
|
v_mov_b32_e32 v129, v47 // 000000003088: 7F02032F
|
|
v_mov_b32_e32 v130, v48 // 00000000308C: 7F040330
|
|
v_mov_b32_e32 v131, v56 // 000000003090: 7F060338
|
|
v_mov_b32_e32 v132, v64 // 000000003094: 7F080340
|
|
v_mov_b32_e32 v133, v49 // 000000003098: 7F0A0331
|
|
v_mov_b32_e32 v134, v57 // 00000000309C: 7F0C0339
|
|
v_mov_b32_e32 v135, v65 // 0000000030A0: 7F0E0341
|
|
v_mov_b32_e32 v136, v50 // 0000000030A4: 7F100332
|
|
v_mov_b32_e32 v137, v58 // 0000000030A8: 7F12033A
|
|
s_waitcnt lgkmcnt(4) // 0000000030AC: BF89FC47
|
|
v_mul_f32_e32 v82, v139, v82 // 0000000030B0: 10A4A58B
|
|
v_add_f32_e32 v76, v138, v82 // 0000000030B4: 0698A58A
|
|
v_mov_b32_e32 v82, v76 // 0000000030B8: 7EA4034C
|
|
v_cvt_f16_f32_e32 v82, v82 // 0000000030BC: 7EA41552
|
|
buffer_store_b16 v82, v79, s[16:19], 0 offen // 0000000030C0: E0640000 8044524F
|
|
s_waitcnt lgkmcnt(2) // 0000000030C8: BF89FC27
|
|
v_mul_f32_e32 v83, v141, v83 // 0000000030CC: 10A6A78D
|
|
v_add_f32_e32 v76, v140, v83 // 0000000030D0: 0698A78C
|
|
v_mov_b32_e32 v83, v76 // 0000000030D4: 7EA6034C
|
|
v_cvt_f16_f32_e32 v83, v83 // 0000000030D8: 7EA61553
|
|
buffer_store_b16 v83, v79, s[16:19], 0 offen offset:64 // 0000000030DC: E0640040 8044534F
|
|
s_waitcnt lgkmcnt(0) // 0000000030E4: BF89FC07
|
|
v_mul_f32_e32 v84, v143, v84 // 0000000030E8: 10A8A98F
|
|
v_add_f32_e32 v76, v142, v84 // 0000000030EC: 0698A98E
|
|
v_mov_b32_e32 v84, v76 // 0000000030F0: 7EA8034C
|
|
v_cvt_f16_f32_e32 v84, v84 // 0000000030F4: 7EA81554
|
|
buffer_store_b16 v84, v79, s[16:19], 0 offen offset:128 // 0000000030F8: E0640080 8044544F
|
|
v_mul_f32_e32 v85, v139, v85 // 000000003100: 10AAAB8B
|
|
v_add_f32_e32 v76, v138, v85 // 000000003104: 0698AB8A
|
|
v_mov_b32_e32 v85, v76 // 000000003108: 7EAA034C
|
|
v_cvt_f16_f32_e32 v85, v85 // 00000000310C: 7EAA1555
|
|
s_mul_i32 s8, s36, 4 // 000000003110: 96088424
|
|
s_add_u32 s16, s16, s8 // 000000003114: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003118: 82118011
|
|
buffer_store_b16 v85, v79, s[16:19], 0 offen // 00000000311C: E0640000 8044554F
|
|
v_mul_f32_e32 v86, v141, v86 // 000000003124: 10ACAD8D
|
|
v_add_f32_e32 v76, v140, v86 // 000000003128: 0698AD8C
|
|
v_mov_b32_e32 v86, v76 // 00000000312C: 7EAC034C
|
|
v_cvt_f16_f32_e32 v86, v86 // 000000003130: 7EAC1556
|
|
buffer_store_b16 v86, v79, s[16:19], 0 offen offset:64 // 000000003134: E0640040 8044564F
|
|
v_mul_f32_e32 v87, v143, v87 // 00000000313C: 10AEAF8F
|
|
v_add_f32_e32 v76, v142, v87 // 000000003140: 0698AF8E
|
|
v_mov_b32_e32 v87, v76 // 000000003144: 7EAE034C
|
|
v_cvt_f16_f32_e32 v87, v87 // 000000003148: 7EAE1557
|
|
buffer_store_b16 v87, v79, s[16:19], 0 offen offset:128 // 00000000314C: E0640080 8044574F
|
|
v_mul_f32_e32 v88, v139, v88 // 000000003154: 10B0B18B
|
|
v_add_f32_e32 v76, v138, v88 // 000000003158: 0698B18A
|
|
v_mov_b32_e32 v88, v76 // 00000000315C: 7EB0034C
|
|
v_cvt_f16_f32_e32 v88, v88 // 000000003160: 7EB01558
|
|
s_mul_i32 s8, s36, 4 // 000000003164: 96088424
|
|
s_add_u32 s16, s16, s8 // 000000003168: 80100810
|
|
s_addc_u32 s17, s17, 0 // 00000000316C: 82118011
|
|
buffer_store_b16 v88, v79, s[16:19], 0 offen // 000000003170: E0640000 8044584F
|
|
v_mul_f32_e32 v89, v141, v89 // 000000003178: 10B2B38D
|
|
v_add_f32_e32 v76, v140, v89 // 00000000317C: 0698B38C
|
|
v_mov_b32_e32 v89, v76 // 000000003180: 7EB2034C
|
|
v_cvt_f16_f32_e32 v89, v89 // 000000003184: 7EB21559
|
|
buffer_store_b16 v89, v79, s[16:19], 0 offen offset:64 // 000000003188: E0640040 8044594F
|
|
v_mul_f32_e32 v90, v143, v90 // 000000003190: 10B4B58F
|
|
v_add_f32_e32 v76, v142, v90 // 000000003194: 0698B58E
|
|
v_mov_b32_e32 v90, v76 // 000000003198: 7EB4034C
|
|
v_cvt_f16_f32_e32 v90, v90 // 00000000319C: 7EB4155A
|
|
buffer_store_b16 v90, v79, s[16:19], 0 offen offset:128 // 0000000031A0: E0640080 80445A4F
|
|
v_mul_f32_e32 v91, v139, v91 // 0000000031A8: 10B6B78B
|
|
v_add_f32_e32 v76, v138, v91 // 0000000031AC: 0698B78A
|
|
v_mov_b32_e32 v91, v76 // 0000000031B0: 7EB6034C
|
|
v_cvt_f16_f32_e32 v91, v91 // 0000000031B4: 7EB6155B
|
|
s_mul_i32 s8, s36, 4 // 0000000031B8: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000031BC: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000031C0: 82118011
|
|
buffer_store_b16 v91, v79, s[16:19], 0 offen // 0000000031C4: E0640000 80445B4F
|
|
v_mul_f32_e32 v92, v141, v92 // 0000000031CC: 10B8B98D
|
|
v_add_f32_e32 v76, v140, v92 // 0000000031D0: 0698B98C
|
|
v_mov_b32_e32 v92, v76 // 0000000031D4: 7EB8034C
|
|
v_cvt_f16_f32_e32 v92, v92 // 0000000031D8: 7EB8155C
|
|
buffer_store_b16 v92, v79, s[16:19], 0 offen offset:64 // 0000000031DC: E0640040 80445C4F
|
|
v_mul_f32_e32 v93, v143, v93 // 0000000031E4: 10BABB8F
|
|
v_add_f32_e32 v76, v142, v93 // 0000000031E8: 0698BB8E
|
|
v_mov_b32_e32 v93, v76 // 0000000031EC: 7EBA034C
|
|
v_cvt_f16_f32_e32 v93, v93 // 0000000031F0: 7EBA155D
|
|
buffer_store_b16 v93, v79, s[16:19], 0 offen offset:128 // 0000000031F4: E0640080 80445D4F
|
|
v_mul_f32_e32 v94, v139, v94 // 0000000031FC: 10BCBD8B
|
|
v_add_f32_e32 v76, v138, v94 // 000000003200: 0698BD8A
|
|
v_mov_b32_e32 v94, v76 // 000000003204: 7EBC034C
|
|
v_cvt_f16_f32_e32 v94, v94 // 000000003208: 7EBC155E
|
|
s_mul_i32 s8, s36, 4 // 00000000320C: 96088424
|
|
s_add_u32 s16, s16, s8 // 000000003210: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003214: 82118011
|
|
buffer_store_b16 v94, v79, s[16:19], 0 offen // 000000003218: E0640000 80445E4F
|
|
v_mul_f32_e32 v95, v141, v95 // 000000003220: 10BEBF8D
|
|
v_add_f32_e32 v76, v140, v95 // 000000003224: 0698BF8C
|
|
v_mov_b32_e32 v95, v76 // 000000003228: 7EBE034C
|
|
v_cvt_f16_f32_e32 v95, v95 // 00000000322C: 7EBE155F
|
|
buffer_store_b16 v95, v79, s[16:19], 0 offen offset:64 // 000000003230: E0640040 80445F4F
|
|
v_mul_f32_e32 v96, v143, v96 // 000000003238: 10C0C18F
|
|
v_add_f32_e32 v76, v142, v96 // 00000000323C: 0698C18E
|
|
v_mov_b32_e32 v96, v76 // 000000003240: 7EC0034C
|
|
v_cvt_f16_f32_e32 v96, v96 // 000000003244: 7EC01560
|
|
buffer_store_b16 v96, v79, s[16:19], 0 offen offset:128 // 000000003248: E0640080 8044604F
|
|
v_mul_f32_e32 v97, v139, v97 // 000000003250: 10C2C38B
|
|
v_add_f32_e32 v76, v138, v97 // 000000003254: 0698C38A
|
|
v_mov_b32_e32 v97, v76 // 000000003258: 7EC2034C
|
|
v_cvt_f16_f32_e32 v97, v97 // 00000000325C: 7EC21561
|
|
s_mul_i32 s8, s36, 4 // 000000003260: 96088424
|
|
s_add_u32 s16, s16, s8 // 000000003264: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003268: 82118011
|
|
buffer_store_b16 v97, v79, s[16:19], 0 offen // 00000000326C: E0640000 8044614F
|
|
v_mul_f32_e32 v98, v141, v98 // 000000003274: 10C4C58D
|
|
v_add_f32_e32 v76, v140, v98 // 000000003278: 0698C58C
|
|
v_mov_b32_e32 v98, v76 // 00000000327C: 7EC4034C
|
|
v_cvt_f16_f32_e32 v98, v98 // 000000003280: 7EC41562
|
|
buffer_store_b16 v98, v79, s[16:19], 0 offen offset:64 // 000000003284: E0640040 8044624F
|
|
v_mul_f32_e32 v99, v143, v99 // 00000000328C: 10C6C78F
|
|
v_add_f32_e32 v76, v142, v99 // 000000003290: 0698C78E
|
|
v_mov_b32_e32 v99, v76 // 000000003294: 7EC6034C
|
|
v_cvt_f16_f32_e32 v99, v99 // 000000003298: 7EC61563
|
|
buffer_store_b16 v99, v79, s[16:19], 0 offen offset:128 // 00000000329C: E0640080 8044634F
|
|
v_mul_f32_e32 v100, v139, v100 // 0000000032A4: 10C8C98B
|
|
v_add_f32_e32 v76, v138, v100 // 0000000032A8: 0698C98A
|
|
v_mov_b32_e32 v100, v76 // 0000000032AC: 7EC8034C
|
|
v_cvt_f16_f32_e32 v100, v100 // 0000000032B0: 7EC81564
|
|
s_mul_i32 s8, s36, 4 // 0000000032B4: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000032B8: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000032BC: 82118011
|
|
buffer_store_b16 v100, v79, s[16:19], 0 offen // 0000000032C0: E0640000 8044644F
|
|
v_mul_f32_e32 v101, v141, v101 // 0000000032C8: 10CACB8D
|
|
v_add_f32_e32 v76, v140, v101 // 0000000032CC: 0698CB8C
|
|
v_mov_b32_e32 v101, v76 // 0000000032D0: 7ECA034C
|
|
v_cvt_f16_f32_e32 v101, v101 // 0000000032D4: 7ECA1565
|
|
buffer_store_b16 v101, v79, s[16:19], 0 offen offset:64 // 0000000032D8: E0640040 8044654F
|
|
v_mul_f32_e32 v102, v143, v102 // 0000000032E0: 10CCCD8F
|
|
v_add_f32_e32 v76, v142, v102 // 0000000032E4: 0698CD8E
|
|
v_mov_b32_e32 v102, v76 // 0000000032E8: 7ECC034C
|
|
v_cvt_f16_f32_e32 v102, v102 // 0000000032EC: 7ECC1566
|
|
buffer_store_b16 v102, v79, s[16:19], 0 offen offset:128 // 0000000032F0: E0640080 8044664F
|
|
v_mul_f32_e32 v103, v139, v103 // 0000000032F8: 10CECF8B
|
|
v_add_f32_e32 v76, v138, v103 // 0000000032FC: 0698CF8A
|
|
v_mov_b32_e32 v103, v76 // 000000003300: 7ECE034C
|
|
v_cvt_f16_f32_e32 v103, v103 // 000000003304: 7ECE1567
|
|
s_mul_i32 s8, s36, 4 // 000000003308: 96088424
|
|
s_add_u32 s16, s16, s8 // 00000000330C: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003310: 82118011
|
|
buffer_store_b16 v103, v79, s[16:19], 0 offen // 000000003314: E0640000 8044674F
|
|
v_mul_f32_e32 v104, v141, v104 // 00000000331C: 10D0D18D
|
|
v_add_f32_e32 v76, v140, v104 // 000000003320: 0698D18C
|
|
v_mov_b32_e32 v104, v76 // 000000003324: 7ED0034C
|
|
v_cvt_f16_f32_e32 v104, v104 // 000000003328: 7ED01568
|
|
buffer_store_b16 v104, v79, s[16:19], 0 offen offset:64 // 00000000332C: E0640040 8044684F
|
|
v_mul_f32_e32 v105, v143, v105 // 000000003334: 10D2D38F
|
|
v_add_f32_e32 v76, v142, v105 // 000000003338: 0698D38E
|
|
v_mov_b32_e32 v105, v76 // 00000000333C: 7ED2034C
|
|
v_cvt_f16_f32_e32 v105, v105 // 000000003340: 7ED21569
|
|
buffer_store_b16 v105, v79, s[16:19], 0 offen offset:128 // 000000003344: E0640080 8044694F
|
|
v_mul_f32_e32 v106, v139, v106 // 00000000334C: 10D4D58B
|
|
v_add_f32_e32 v76, v138, v106 // 000000003350: 0698D58A
|
|
v_mov_b32_e32 v106, v76 // 000000003354: 7ED4034C
|
|
v_cvt_f16_f32_e32 v106, v106 // 000000003358: 7ED4156A
|
|
s_mul_i32 s8, s36, 36 // 00000000335C: 9608A424
|
|
s_add_u32 s16, s16, s8 // 000000003360: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003364: 82118011
|
|
buffer_store_b16 v106, v79, s[16:19], 0 offen // 000000003368: E0640000 80446A4F
|
|
v_mul_f32_e32 v107, v141, v107 // 000000003370: 10D6D78D
|
|
v_add_f32_e32 v76, v140, v107 // 000000003374: 0698D78C
|
|
v_mov_b32_e32 v107, v76 // 000000003378: 7ED6034C
|
|
v_cvt_f16_f32_e32 v107, v107 // 00000000337C: 7ED6156B
|
|
buffer_store_b16 v107, v79, s[16:19], 0 offen offset:64 // 000000003380: E0640040 80446B4F
|
|
v_mul_f32_e32 v108, v143, v108 // 000000003388: 10D8D98F
|
|
v_add_f32_e32 v76, v142, v108 // 00000000338C: 0698D98E
|
|
v_mov_b32_e32 v108, v76 // 000000003390: 7ED8034C
|
|
v_cvt_f16_f32_e32 v108, v108 // 000000003394: 7ED8156C
|
|
buffer_store_b16 v108, v79, s[16:19], 0 offen offset:128 // 000000003398: E0640080 80446C4F
|
|
v_mul_f32_e32 v109, v139, v109 // 0000000033A0: 10DADB8B
|
|
v_add_f32_e32 v76, v138, v109 // 0000000033A4: 0698DB8A
|
|
v_mov_b32_e32 v109, v76 // 0000000033A8: 7EDA034C
|
|
v_cvt_f16_f32_e32 v109, v109 // 0000000033AC: 7EDA156D
|
|
s_mul_i32 s8, s36, 4 // 0000000033B0: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000033B4: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000033B8: 82118011
|
|
buffer_store_b16 v109, v79, s[16:19], 0 offen // 0000000033BC: E0640000 80446D4F
|
|
v_mul_f32_e32 v110, v141, v110 // 0000000033C4: 10DCDD8D
|
|
v_add_f32_e32 v76, v140, v110 // 0000000033C8: 0698DD8C
|
|
v_mov_b32_e32 v110, v76 // 0000000033CC: 7EDC034C
|
|
v_cvt_f16_f32_e32 v110, v110 // 0000000033D0: 7EDC156E
|
|
buffer_store_b16 v110, v79, s[16:19], 0 offen offset:64 // 0000000033D4: E0640040 80446E4F
|
|
v_mul_f32_e32 v111, v143, v111 // 0000000033DC: 10DEDF8F
|
|
v_add_f32_e32 v76, v142, v111 // 0000000033E0: 0698DF8E
|
|
v_mov_b32_e32 v111, v76 // 0000000033E4: 7EDE034C
|
|
v_cvt_f16_f32_e32 v111, v111 // 0000000033E8: 7EDE156F
|
|
buffer_store_b16 v111, v79, s[16:19], 0 offen offset:128 // 0000000033EC: E0640080 80446F4F
|
|
v_mul_f32_e32 v112, v139, v112 // 0000000033F4: 10E0E18B
|
|
v_add_f32_e32 v76, v138, v112 // 0000000033F8: 0698E18A
|
|
v_mov_b32_e32 v112, v76 // 0000000033FC: 7EE0034C
|
|
v_cvt_f16_f32_e32 v112, v112 // 000000003400: 7EE01570
|
|
s_mul_i32 s8, s36, 4 // 000000003404: 96088424
|
|
s_add_u32 s16, s16, s8 // 000000003408: 80100810
|
|
s_addc_u32 s17, s17, 0 // 00000000340C: 82118011
|
|
buffer_store_b16 v112, v79, s[16:19], 0 offen // 000000003410: E0640000 8044704F
|
|
v_mul_f32_e32 v113, v141, v113 // 000000003418: 10E2E38D
|
|
v_add_f32_e32 v76, v140, v113 // 00000000341C: 0698E38C
|
|
v_mov_b32_e32 v113, v76 // 000000003420: 7EE2034C
|
|
v_cvt_f16_f32_e32 v113, v113 // 000000003424: 7EE21571
|
|
buffer_store_b16 v113, v79, s[16:19], 0 offen offset:64 // 000000003428: E0640040 8044714F
|
|
v_mul_f32_e32 v114, v143, v114 // 000000003430: 10E4E58F
|
|
v_add_f32_e32 v76, v142, v114 // 000000003434: 0698E58E
|
|
v_mov_b32_e32 v114, v76 // 000000003438: 7EE4034C
|
|
v_cvt_f16_f32_e32 v114, v114 // 00000000343C: 7EE41572
|
|
buffer_store_b16 v114, v79, s[16:19], 0 offen offset:128 // 000000003440: E0640080 8044724F
|
|
v_mul_f32_e32 v115, v139, v115 // 000000003448: 10E6E78B
|
|
v_add_f32_e32 v76, v138, v115 // 00000000344C: 0698E78A
|
|
v_mov_b32_e32 v115, v76 // 000000003450: 7EE6034C
|
|
v_cvt_f16_f32_e32 v115, v115 // 000000003454: 7EE61573
|
|
s_mul_i32 s8, s36, 4 // 000000003458: 96088424
|
|
s_add_u32 s16, s16, s8 // 00000000345C: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003460: 82118011
|
|
buffer_store_b16 v115, v79, s[16:19], 0 offen // 000000003464: E0640000 8044734F
|
|
v_mul_f32_e32 v116, v141, v116 // 00000000346C: 10E8E98D
|
|
v_add_f32_e32 v76, v140, v116 // 000000003470: 0698E98C
|
|
v_mov_b32_e32 v116, v76 // 000000003474: 7EE8034C
|
|
v_cvt_f16_f32_e32 v116, v116 // 000000003478: 7EE81574
|
|
buffer_store_b16 v116, v79, s[16:19], 0 offen offset:64 // 00000000347C: E0640040 8044744F
|
|
v_mul_f32_e32 v117, v143, v117 // 000000003484: 10EAEB8F
|
|
v_add_f32_e32 v76, v142, v117 // 000000003488: 0698EB8E
|
|
v_mov_b32_e32 v117, v76 // 00000000348C: 7EEA034C
|
|
v_cvt_f16_f32_e32 v117, v117 // 000000003490: 7EEA1575
|
|
buffer_store_b16 v117, v79, s[16:19], 0 offen offset:128 // 000000003494: E0640080 8044754F
|
|
v_mul_f32_e32 v118, v139, v118 // 00000000349C: 10ECED8B
|
|
v_add_f32_e32 v76, v138, v118 // 0000000034A0: 0698ED8A
|
|
v_mov_b32_e32 v118, v76 // 0000000034A4: 7EEC034C
|
|
v_cvt_f16_f32_e32 v118, v118 // 0000000034A8: 7EEC1576
|
|
s_mul_i32 s8, s36, 4 // 0000000034AC: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000034B0: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000034B4: 82118011
|
|
buffer_store_b16 v118, v79, s[16:19], 0 offen // 0000000034B8: E0640000 8044764F
|
|
v_mul_f32_e32 v119, v141, v119 // 0000000034C0: 10EEEF8D
|
|
v_add_f32_e32 v76, v140, v119 // 0000000034C4: 0698EF8C
|
|
v_mov_b32_e32 v119, v76 // 0000000034C8: 7EEE034C
|
|
v_cvt_f16_f32_e32 v119, v119 // 0000000034CC: 7EEE1577
|
|
buffer_store_b16 v119, v79, s[16:19], 0 offen offset:64 // 0000000034D0: E0640040 8044774F
|
|
v_mul_f32_e32 v120, v143, v120 // 0000000034D8: 10F0F18F
|
|
v_add_f32_e32 v76, v142, v120 // 0000000034DC: 0698F18E
|
|
v_mov_b32_e32 v120, v76 // 0000000034E0: 7EF0034C
|
|
v_cvt_f16_f32_e32 v120, v120 // 0000000034E4: 7EF01578
|
|
buffer_store_b16 v120, v79, s[16:19], 0 offen offset:128 // 0000000034E8: E0640080 8044784F
|
|
v_mul_f32_e32 v121, v139, v121 // 0000000034F0: 10F2F38B
|
|
v_add_f32_e32 v76, v138, v121 // 0000000034F4: 0698F38A
|
|
v_mov_b32_e32 v121, v76 // 0000000034F8: 7EF2034C
|
|
v_cvt_f16_f32_e32 v121, v121 // 0000000034FC: 7EF21579
|
|
s_mul_i32 s8, s36, 4 // 000000003500: 96088424
|
|
s_add_u32 s16, s16, s8 // 000000003504: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003508: 82118011
|
|
buffer_store_b16 v121, v79, s[16:19], 0 offen // 00000000350C: E0640000 8044794F
|
|
v_mul_f32_e32 v122, v141, v122 // 000000003514: 10F4F58D
|
|
v_add_f32_e32 v76, v140, v122 // 000000003518: 0698F58C
|
|
v_mov_b32_e32 v122, v76 // 00000000351C: 7EF4034C
|
|
v_cvt_f16_f32_e32 v122, v122 // 000000003520: 7EF4157A
|
|
buffer_store_b16 v122, v79, s[16:19], 0 offen offset:64 // 000000003524: E0640040 80447A4F
|
|
v_mul_f32_e32 v123, v143, v123 // 00000000352C: 10F6F78F
|
|
v_add_f32_e32 v76, v142, v123 // 000000003530: 0698F78E
|
|
v_mov_b32_e32 v123, v76 // 000000003534: 7EF6034C
|
|
v_cvt_f16_f32_e32 v123, v123 // 000000003538: 7EF6157B
|
|
buffer_store_b16 v123, v79, s[16:19], 0 offen offset:128 // 00000000353C: E0640080 80447B4F
|
|
v_mul_f32_e32 v124, v139, v124 // 000000003544: 10F8F98B
|
|
v_add_f32_e32 v76, v138, v124 // 000000003548: 0698F98A
|
|
v_mov_b32_e32 v124, v76 // 00000000354C: 7EF8034C
|
|
v_cvt_f16_f32_e32 v124, v124 // 000000003550: 7EF8157C
|
|
s_mul_i32 s8, s36, 4 // 000000003554: 96088424
|
|
s_add_u32 s16, s16, s8 // 000000003558: 80100810
|
|
s_addc_u32 s17, s17, 0 // 00000000355C: 82118011
|
|
buffer_store_b16 v124, v79, s[16:19], 0 offen // 000000003560: E0640000 80447C4F
|
|
v_mul_f32_e32 v125, v141, v125 // 000000003568: 10FAFB8D
|
|
v_add_f32_e32 v76, v140, v125 // 00000000356C: 0698FB8C
|
|
v_mov_b32_e32 v125, v76 // 000000003570: 7EFA034C
|
|
v_cvt_f16_f32_e32 v125, v125 // 000000003574: 7EFA157D
|
|
buffer_store_b16 v125, v79, s[16:19], 0 offen offset:64 // 000000003578: E0640040 80447D4F
|
|
v_mul_f32_e32 v126, v143, v126 // 000000003580: 10FCFD8F
|
|
v_add_f32_e32 v76, v142, v126 // 000000003584: 0698FD8E
|
|
v_mov_b32_e32 v126, v76 // 000000003588: 7EFC034C
|
|
v_cvt_f16_f32_e32 v126, v126 // 00000000358C: 7EFC157E
|
|
buffer_store_b16 v126, v79, s[16:19], 0 offen offset:128 // 000000003590: E0640080 80447E4F
|
|
v_mul_f32_e32 v127, v139, v127 // 000000003598: 10FEFF8B
|
|
v_add_f32_e32 v76, v138, v127 // 00000000359C: 0698FF8A
|
|
v_mov_b32_e32 v127, v76 // 0000000035A0: 7EFE034C
|
|
v_cvt_f16_f32_e32 v127, v127 // 0000000035A4: 7EFE157F
|
|
s_mul_i32 s8, s36, 4 // 0000000035A8: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000035AC: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000035B0: 82118011
|
|
buffer_store_b16 v127, v79, s[16:19], 0 offen // 0000000035B4: E0640000 80447F4F
|
|
v_mul_f32_e32 v128, v141, v128 // 0000000035BC: 1101018D
|
|
v_add_f32_e32 v76, v140, v128 // 0000000035C0: 0699018C
|
|
v_mov_b32_e32 v128, v76 // 0000000035C4: 7F00034C
|
|
v_cvt_f16_f32_e64 v128, v128 // 0000000035C8: D58A0080 00000180
|
|
buffer_store_b16 v128, v79, s[16:19], 0 offen offset:64 // 0000000035D0: E0640040 8044804F
|
|
v_mul_f32_e32 v129, v143, v129 // 0000000035D8: 1103038F
|
|
v_add_f32_e32 v76, v142, v129 // 0000000035DC: 0699038E
|
|
v_mov_b32_e32 v129, v76 // 0000000035E0: 7F02034C
|
|
v_cvt_f16_f32_e64 v129, v129 // 0000000035E4: D58A0081 00000181
|
|
buffer_store_b16 v129, v79, s[16:19], 0 offen offset:128 // 0000000035EC: E0640080 8044814F
|
|
v_mul_f32_e32 v130, v139, v130 // 0000000035F4: 1105058B
|
|
v_add_f32_e32 v76, v138, v130 // 0000000035F8: 0699058A
|
|
v_mov_b32_e32 v130, v76 // 0000000035FC: 7F04034C
|
|
v_cvt_f16_f32_e64 v130, v130 // 000000003600: D58A0082 00000182
|
|
s_mul_i32 s8, s36, 36 // 000000003608: 9608A424
|
|
s_add_u32 s16, s16, s8 // 00000000360C: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003610: 82118011
|
|
buffer_store_b16 v130, v79, s[16:19], 0 offen // 000000003614: E0640000 8044824F
|
|
v_mul_f32_e32 v131, v141, v131 // 00000000361C: 1107078D
|
|
v_add_f32_e32 v76, v140, v131 // 000000003620: 0699078C
|
|
v_mov_b32_e32 v131, v76 // 000000003624: 7F06034C
|
|
v_cvt_f16_f32_e64 v131, v131 // 000000003628: D58A0083 00000183
|
|
buffer_store_b16 v131, v79, s[16:19], 0 offen offset:64 // 000000003630: E0640040 8044834F
|
|
v_mul_f32_e32 v132, v143, v132 // 000000003638: 1109098F
|
|
v_add_f32_e32 v76, v142, v132 // 00000000363C: 0699098E
|
|
v_mov_b32_e32 v132, v76 // 000000003640: 7F08034C
|
|
v_cvt_f16_f32_e64 v132, v132 // 000000003644: D58A0084 00000184
|
|
buffer_store_b16 v132, v79, s[16:19], 0 offen offset:128 // 00000000364C: E0640080 8044844F
|
|
v_mul_f32_e32 v133, v139, v133 // 000000003654: 110B0B8B
|
|
v_add_f32_e32 v76, v138, v133 // 000000003658: 06990B8A
|
|
v_mov_b32_e32 v133, v76 // 00000000365C: 7F0A034C
|
|
v_cvt_f16_f32_e64 v133, v133 // 000000003660: D58A0085 00000185
|
|
s_mul_i32 s8, s36, 4 // 000000003668: 96088424
|
|
s_add_u32 s16, s16, s8 // 00000000366C: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003670: 82118011
|
|
buffer_store_b16 v133, v79, s[16:19], 0 offen // 000000003674: E0640000 8044854F
|
|
v_mul_f32_e32 v134, v141, v134 // 00000000367C: 110D0D8D
|
|
v_add_f32_e32 v76, v140, v134 // 000000003680: 06990D8C
|
|
v_mov_b32_e32 v134, v76 // 000000003684: 7F0C034C
|
|
v_cvt_f16_f32_e64 v134, v134 // 000000003688: D58A0086 00000186
|
|
buffer_store_b16 v134, v79, s[16:19], 0 offen offset:64 // 000000003690: E0640040 8044864F
|
|
v_mul_f32_e32 v135, v143, v135 // 000000003698: 110F0F8F
|
|
v_add_f32_e32 v76, v142, v135 // 00000000369C: 06990F8E
|
|
v_mov_b32_e32 v135, v76 // 0000000036A0: 7F0E034C
|
|
v_cvt_f16_f32_e64 v135, v135 // 0000000036A4: D58A0087 00000187
|
|
buffer_store_b16 v135, v79, s[16:19], 0 offen offset:128 // 0000000036AC: E0640080 8044874F
|
|
v_mul_f32_e32 v136, v139, v136 // 0000000036B4: 1111118B
|
|
v_add_f32_e32 v76, v138, v136 // 0000000036B8: 0699118A
|
|
v_mov_b32_e32 v136, v76 // 0000000036BC: 7F10034C
|
|
v_cvt_f16_f32_e64 v136, v136 // 0000000036C0: D58A0088 00000188
|
|
s_mul_i32 s8, s36, 4 // 0000000036C8: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000036CC: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000036D0: 82118011
|
|
buffer_store_b16 v136, v79, s[16:19], 0 offen // 0000000036D4: E0640000 8044884F
|
|
v_mul_f32_e32 v137, v141, v137 // 0000000036DC: 1113138D
|
|
v_add_f32_e32 v76, v140, v137 // 0000000036E0: 0699138C
|
|
v_mov_b32_e32 v137, v76 // 0000000036E4: 7F12034C
|
|
v_cvt_f16_f32_e64 v137, v137 // 0000000036E8: D58A0089 00000189
|
|
buffer_store_b16 v137, v79, s[16:19], 0 offen offset:64 // 0000000036F0: E0640040 8044894F
|
|
s_nop 0 // 0000000036F8: BF800000
|
|
ds_load_b32 v98, v81 offset:256 // 0000000036FC: D8D80100 62000051
|
|
ds_load_b32 v99, v81 offset:768 // 000000003704: D8D80300 63000051
|
|
ds_load_b32 v100, v81 // 00000000370C: D8D80000 64000051
|
|
ds_load_b32 v101, v81 offset:512 // 000000003714: D8D80200 65000051
|
|
ds_load_b32 v102, v81 offset:128 // 00000000371C: D8D80080 66000051
|
|
ds_load_b32 v103, v81 offset:640 // 000000003724: D8D80280 67000051
|
|
v_mov_b32_e32 v82, v66 // 00000000372C: 7EA40342
|
|
v_mov_b32_e32 v83, v51 // 000000003730: 7EA60333
|
|
v_mov_b32_e32 v84, v59 // 000000003734: 7EA8033B
|
|
v_mov_b32_e32 v85, v67 // 000000003738: 7EAA0343
|
|
v_mov_b32_e32 v86, v52 // 00000000373C: 7EAC0334
|
|
v_mov_b32_e32 v87, v60 // 000000003740: 7EAE033C
|
|
v_mov_b32_e32 v88, v68 // 000000003744: 7EB00344
|
|
v_mov_b32_e32 v89, v53 // 000000003748: 7EB20335
|
|
v_mov_b32_e32 v90, v61 // 00000000374C: 7EB4033D
|
|
v_mov_b32_e32 v91, v69 // 000000003750: 7EB60345
|
|
v_mov_b32_e32 v92, v54 // 000000003754: 7EB80336
|
|
v_mov_b32_e32 v93, v62 // 000000003758: 7EBA033E
|
|
v_mov_b32_e32 v94, v70 // 00000000375C: 7EBC0346
|
|
v_mov_b32_e32 v95, v55 // 000000003760: 7EBE0337
|
|
v_mov_b32_e32 v96, v63 // 000000003764: 7EC0033F
|
|
v_mov_b32_e32 v97, v71 // 000000003768: 7EC20347
|
|
s_waitcnt lgkmcnt(4) // 00000000376C: BF89FC47
|
|
v_mul_f32_e32 v82, v99, v82 // 000000003770: 10A4A563
|
|
v_add_f32_e32 v76, v98, v82 // 000000003774: 0698A562
|
|
v_mov_b32_e32 v82, v76 // 000000003778: 7EA4034C
|
|
v_cvt_f16_f32_e32 v82, v82 // 00000000377C: 7EA41552
|
|
buffer_store_b16 v82, v79, s[16:19], 0 offen offset:128 // 000000003780: E0640080 8044524F
|
|
s_waitcnt lgkmcnt(2) // 000000003788: BF89FC27
|
|
v_mul_f32_e32 v83, v101, v83 // 00000000378C: 10A6A765
|
|
v_add_f32_e32 v76, v100, v83 // 000000003790: 0698A764
|
|
v_mov_b32_e32 v83, v76 // 000000003794: 7EA6034C
|
|
v_cvt_f16_f32_e32 v83, v83 // 000000003798: 7EA61553
|
|
s_mul_i32 s8, s36, 4 // 00000000379C: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000037A0: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000037A4: 82118011
|
|
buffer_store_b16 v83, v79, s[16:19], 0 offen // 0000000037A8: E0640000 8044534F
|
|
s_waitcnt lgkmcnt(0) // 0000000037B0: BF89FC07
|
|
v_mul_f32_e32 v84, v103, v84 // 0000000037B4: 10A8A967
|
|
v_add_f32_e32 v76, v102, v84 // 0000000037B8: 0698A966
|
|
v_mov_b32_e32 v84, v76 // 0000000037BC: 7EA8034C
|
|
v_cvt_f16_f32_e32 v84, v84 // 0000000037C0: 7EA81554
|
|
buffer_store_b16 v84, v79, s[16:19], 0 offen offset:64 // 0000000037C4: E0640040 8044544F
|
|
v_mul_f32_e32 v85, v99, v85 // 0000000037CC: 10AAAB63
|
|
v_add_f32_e32 v76, v98, v85 // 0000000037D0: 0698AB62
|
|
v_mov_b32_e32 v85, v76 // 0000000037D4: 7EAA034C
|
|
v_cvt_f16_f32_e32 v85, v85 // 0000000037D8: 7EAA1555
|
|
buffer_store_b16 v85, v79, s[16:19], 0 offen offset:128 // 0000000037DC: E0640080 8044554F
|
|
v_mul_f32_e32 v86, v101, v86 // 0000000037E4: 10ACAD65
|
|
v_add_f32_e32 v76, v100, v86 // 0000000037E8: 0698AD64
|
|
v_mov_b32_e32 v86, v76 // 0000000037EC: 7EAC034C
|
|
v_cvt_f16_f32_e32 v86, v86 // 0000000037F0: 7EAC1556
|
|
s_mul_i32 s8, s36, 4 // 0000000037F4: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000037F8: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000037FC: 82118011
|
|
buffer_store_b16 v86, v79, s[16:19], 0 offen // 000000003800: E0640000 8044564F
|
|
v_mul_f32_e32 v87, v103, v87 // 000000003808: 10AEAF67
|
|
v_add_f32_e32 v76, v102, v87 // 00000000380C: 0698AF66
|
|
v_mov_b32_e32 v87, v76 // 000000003810: 7EAE034C
|
|
v_cvt_f16_f32_e32 v87, v87 // 000000003814: 7EAE1557
|
|
buffer_store_b16 v87, v79, s[16:19], 0 offen offset:64 // 000000003818: E0640040 8044574F
|
|
v_mul_f32_e32 v88, v99, v88 // 000000003820: 10B0B163
|
|
v_add_f32_e32 v76, v98, v88 // 000000003824: 0698B162
|
|
v_mov_b32_e32 v88, v76 // 000000003828: 7EB0034C
|
|
v_cvt_f16_f32_e32 v88, v88 // 00000000382C: 7EB01558
|
|
buffer_store_b16 v88, v79, s[16:19], 0 offen offset:128 // 000000003830: E0640080 8044584F
|
|
v_mul_f32_e32 v89, v101, v89 // 000000003838: 10B2B365
|
|
v_add_f32_e32 v76, v100, v89 // 00000000383C: 0698B364
|
|
v_mov_b32_e32 v89, v76 // 000000003840: 7EB2034C
|
|
v_cvt_f16_f32_e32 v89, v89 // 000000003844: 7EB21559
|
|
s_mul_i32 s8, s36, 4 // 000000003848: 96088424
|
|
s_add_u32 s16, s16, s8 // 00000000384C: 80100810
|
|
s_addc_u32 s17, s17, 0 // 000000003850: 82118011
|
|
buffer_store_b16 v89, v79, s[16:19], 0 offen // 000000003854: E0640000 8044594F
|
|
v_mul_f32_e32 v90, v103, v90 // 00000000385C: 10B4B567
|
|
v_add_f32_e32 v76, v102, v90 // 000000003860: 0698B566
|
|
v_mov_b32_e32 v90, v76 // 000000003864: 7EB4034C
|
|
v_cvt_f16_f32_e32 v90, v90 // 000000003868: 7EB4155A
|
|
buffer_store_b16 v90, v79, s[16:19], 0 offen offset:64 // 00000000386C: E0640040 80445A4F
|
|
v_mul_f32_e32 v91, v99, v91 // 000000003874: 10B6B763
|
|
v_add_f32_e32 v76, v98, v91 // 000000003878: 0698B762
|
|
v_mov_b32_e32 v91, v76 // 00000000387C: 7EB6034C
|
|
v_cvt_f16_f32_e32 v91, v91 // 000000003880: 7EB6155B
|
|
buffer_store_b16 v91, v79, s[16:19], 0 offen offset:128 // 000000003884: E0640080 80445B4F
|
|
v_mul_f32_e32 v92, v101, v92 // 00000000388C: 10B8B965
|
|
v_add_f32_e32 v76, v100, v92 // 000000003890: 0698B964
|
|
v_mov_b32_e32 v92, v76 // 000000003894: 7EB8034C
|
|
v_cvt_f16_f32_e32 v92, v92 // 000000003898: 7EB8155C
|
|
s_mul_i32 s8, s36, 4 // 00000000389C: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000038A0: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000038A4: 82118011
|
|
buffer_store_b16 v92, v79, s[16:19], 0 offen // 0000000038A8: E0640000 80445C4F
|
|
v_mul_f32_e32 v93, v103, v93 // 0000000038B0: 10BABB67
|
|
v_add_f32_e32 v76, v102, v93 // 0000000038B4: 0698BB66
|
|
v_mov_b32_e32 v93, v76 // 0000000038B8: 7EBA034C
|
|
v_cvt_f16_f32_e32 v93, v93 // 0000000038BC: 7EBA155D
|
|
buffer_store_b16 v93, v79, s[16:19], 0 offen offset:64 // 0000000038C0: E0640040 80445D4F
|
|
v_mul_f32_e32 v94, v99, v94 // 0000000038C8: 10BCBD63
|
|
v_add_f32_e32 v76, v98, v94 // 0000000038CC: 0698BD62
|
|
v_mov_b32_e32 v94, v76 // 0000000038D0: 7EBC034C
|
|
v_cvt_f16_f32_e32 v94, v94 // 0000000038D4: 7EBC155E
|
|
buffer_store_b16 v94, v79, s[16:19], 0 offen offset:128 // 0000000038D8: E0640080 80445E4F
|
|
v_mul_f32_e32 v95, v101, v95 // 0000000038E0: 10BEBF65
|
|
v_add_f32_e32 v76, v100, v95 // 0000000038E4: 0698BF64
|
|
v_mov_b32_e32 v95, v76 // 0000000038E8: 7EBE034C
|
|
v_cvt_f16_f32_e32 v95, v95 // 0000000038EC: 7EBE155F
|
|
s_mul_i32 s8, s36, 4 // 0000000038F0: 96088424
|
|
s_add_u32 s16, s16, s8 // 0000000038F4: 80100810
|
|
s_addc_u32 s17, s17, 0 // 0000000038F8: 82118011
|
|
buffer_store_b16 v95, v79, s[16:19], 0 offen // 0000000038FC: E0640000 80445F4F
|
|
v_mul_f32_e32 v96, v103, v96 // 000000003904: 10C0C167
|
|
v_add_f32_e32 v76, v102, v96 // 000000003908: 0698C166
|
|
v_mov_b32_e32 v96, v76 // 00000000390C: 7EC0034C
|
|
v_cvt_f16_f32_e32 v96, v96 // 000000003910: 7EC01560
|
|
buffer_store_b16 v96, v79, s[16:19], 0 offen offset:64 // 000000003914: E0640040 8044604F
|
|
v_mul_f32_e32 v97, v99, v97 // 00000000391C: 10C2C363
|
|
v_add_f32_e32 v76, v98, v97 // 000000003920: 0698C362
|
|
v_mov_b32_e32 v97, v76 // 000000003924: 7EC2034C
|
|
v_cvt_f16_f32_e32 v97, v97 // 000000003928: 7EC21561
|
|
buffer_store_b16 v97, v79, s[16:19], 0 offen offset:128 // 00000000392C: E0640080 8044614F
|
|
s_nop 0 // 000000003934: BF800000
|
|
s_branch exit // 000000003938: BFA00871
|
|
|
|
epilogue: // 000000000000393c
|
|
s_waitcnt lgkmcnt(4) // 00000000393C: BF89FC47
|
|
v_wmma_f32_16x16x16_f16 v[0:7], v[181:188], v[84:91], v[0:7]// 000000003940: CC404000 1C02A9B5
|
|
ds_load_u16 v108, v80 offset:3104 // 000000003948: D8F00C20 6C000050
|
|
ds_load_u16_d16_hi v108, v80 offset:3296 // 000000003950: DA9C0CE0 6C000050
|
|
ds_load_u16 v109, v80 offset:3488 // 000000003958: D8F00DA0 6D000050
|
|
ds_load_u16_d16_hi v109, v80 offset:3680 // 000000003960: DA9C0E60 6D000050
|
|
ds_load_u16 v110, v80 offset:3872 // 000000003968: D8F00F20 6E000050
|
|
ds_load_u16_d16_hi v110, v80 offset:4064 // 000000003970: DA9C0FE0 6E000050
|
|
ds_load_u16 v111, v80 offset:4256 // 000000003978: D8F010A0 6F000050
|
|
v_wmma_f32_16x16x16_f16 v[8:15], v[181:188], v[92:99], v[8:15]// 000000003980: CC404008 1C22B9B5
|
|
ds_load_u16_d16_hi v111, v80 offset:4448 // 000000003988: DA9C1160 6F000050
|
|
ds_load_u16 v112, v80 offset:4640 // 000000003990: D8F01220 70000050
|
|
ds_load_u16_d16_hi v112, v80 offset:4832 // 000000003998: DA9C12E0 70000050
|
|
ds_load_u16 v113, v80 offset:5024 // 0000000039A0: D8F013A0 71000050
|
|
ds_load_u16_d16_hi v113, v80 offset:5216 // 0000000039A8: DA9C1460 71000050
|
|
ds_load_u16 v114, v80 offset:5408 // 0000000039B0: D8F01520 72000050
|
|
ds_load_u16_d16_hi v114, v80 offset:5600 // 0000000039B8: DA9C15E0 72000050
|
|
v_wmma_f32_16x16x16_f16 v[16:23], v[181:188], v[100:107], v[16:23]// 0000000039C0: CC404010 1C42C9B5
|
|
ds_load_u16 v115, v80 offset:5792 // 0000000039C8: D8F016A0 73000050
|
|
ds_load_u16_d16_hi v115, v80 offset:5984 // 0000000039D0: DA9C1760 73000050
|
|
ds_load_b128 v[205:208], v81 offset:32 // 0000000039D8: DBFC0020 CD000051
|
|
ds_load_b128 v[209:212], v81 offset:48 // 0000000039E0: DBFC0030 D1000051
|
|
ds_load_u16 v116, v80 offset:3168 // 0000000039E8: D8F00C60 74000050
|
|
ds_load_u16_d16_hi v116, v80 offset:3360 // 0000000039F0: DA9C0D20 74000050
|
|
ds_load_u16 v117, v80 offset:3552 // 0000000039F8: D8F00DE0 75000050
|
|
s_waitcnt lgkmcnt(21) // 000000003A00: BF89FD57
|
|
v_wmma_f32_16x16x16_f16 v[24:31], v[189:196], v[84:91], v[24:31]// 000000003A04: CC404018 1C62A9BD
|
|
ds_load_u16_d16_hi v117, v80 offset:3744 // 000000003A0C: DA9C0EA0 75000050
|
|
ds_load_u16 v118, v80 offset:3936 // 000000003A14: D8F00F60 76000050
|
|
ds_load_u16_d16_hi v118, v80 offset:4128 // 000000003A1C: DA9C1020 76000050
|
|
ds_load_u16 v119, v80 offset:4320 // 000000003A24: D8F010E0 77000050
|
|
ds_load_u16_d16_hi v119, v80 offset:4512 // 000000003A2C: DA9C11A0 77000050
|
|
ds_load_u16 v120, v80 offset:4704 // 000000003A34: D8F01260 78000050
|
|
ds_load_u16_d16_hi v120, v80 offset:4896 // 000000003A3C: DA9C1320 78000050
|
|
v_wmma_f32_16x16x16_f16 v[32:39], v[189:196], v[92:99], v[32:39]// 000000003A44: CC404020 1C82B9BD
|
|
ds_load_u16 v121, v80 offset:5088 // 000000003A4C: D8F013E0 79000050
|
|
ds_load_u16_d16_hi v121, v80 offset:5280 // 000000003A54: DA9C14A0 79000050
|
|
ds_load_u16 v122, v80 offset:5472 // 000000003A5C: D8F01560 7A000050
|
|
ds_load_u16_d16_hi v122, v80 offset:5664 // 000000003A64: DA9C1620 7A000050
|
|
ds_load_u16 v123, v80 offset:5856 // 000000003A6C: D8F016E0 7B000050
|
|
ds_load_u16_d16_hi v123, v80 offset:6048 // 000000003A74: DA9C17A0 7B000050
|
|
ds_load_u16 v124, v80 offset:3232 // 000000003A7C: D8F00CA0 7C000050
|
|
v_wmma_f32_16x16x16_f16 v[40:47], v[189:196], v[100:107], v[40:47]// 000000003A84: CC404028 1CA2C9BD
|
|
ds_load_u16_d16_hi v124, v80 offset:3424 // 000000003A8C: DA9C0D60 7C000050
|
|
ds_load_u16 v125, v80 offset:3616 // 000000003A94: D8F00E20 7D000050
|
|
ds_load_u16_d16_hi v125, v80 offset:3808 // 000000003A9C: DA9C0EE0 7D000050
|
|
ds_load_u16 v126, v80 offset:4000 // 000000003AA4: D8F00FA0 7E000050
|
|
ds_load_u16_d16_hi v126, v80 offset:4192 // 000000003AAC: DA9C1060 7E000050
|
|
ds_load_u16 v127, v80 offset:4384 // 000000003AB4: D8F01120 7F000050
|
|
ds_load_u16_d16_hi v127, v80 offset:4576 // 000000003ABC: DA9C11E0 7F000050
|
|
v_wmma_f32_16x16x16_f16 v[48:55], v[197:204], v[84:91], v[48:55]// 000000003AC4: CC404030 1CC2A9C5
|
|
ds_load_u16 v128, v80 offset:4768 // 000000003ACC: D8F012A0 80000050
|
|
ds_load_u16_d16_hi v128, v80 offset:4960 // 000000003AD4: DA9C1360 80000050
|
|
ds_load_u16 v129, v80 offset:5152 // 000000003ADC: D8F01420 81000050
|
|
ds_load_u16_d16_hi v129, v80 offset:5344 // 000000003AE4: DA9C14E0 81000050
|
|
ds_load_u16 v130, v80 offset:5536 // 000000003AEC: D8F015A0 82000050
|
|
ds_load_u16_d16_hi v130, v80 offset:5728 // 000000003AF4: DA9C1660 82000050
|
|
v_wmma_f32_16x16x16_f16 v[56:63], v[197:204], v[92:99], v[56:63]// 000000003AFC: CC404038 1CE2B9C5
|
|
ds_load_u16 v131, v80 offset:5920 // 000000003B04: D8F01720 83000050
|
|
ds_load_u16_d16_hi v131, v80 offset:6112 // 000000003B0C: DA9C17E0 83000050
|
|
ds_load_b128 v[213:216], v81 offset:2592 // 000000003B14: DBFC0A20 D5000051
|
|
ds_load_b128 v[217:220], v81 offset:2608 // 000000003B1C: DBFC0A30 D9000051
|
|
ds_load_b128 v[221:224], v81 offset:5152 // 000000003B24: DBFC1420 DD000051
|
|
ds_load_b128 v[225:228], v81 offset:5168 // 000000003B2C: DBFC1430 E1000051
|
|
v_wmma_f32_16x16x16_f16 v[64:71], v[197:204], v[100:107], v[64:71]// 000000003B34: CC404040 1D02C9C5
|
|
s_waitcnt lgkmcnt(0) // 000000003B3C: BF89FC07
|
|
v_wmma_f32_16x16x16_f16 v[0:7], v[205:212], v[108:115], v[0:7]// 000000003B40: CC404000 1C02D9CD
|
|
v_wmma_f32_16x16x16_f16 v[8:15], v[205:212], v[116:123], v[8:15]// 000000003B48: CC404008 1C22E9CD
|
|
v_wmma_f32_16x16x16_f16 v[16:23], v[205:212], v[124:131], v[16:23]// 000000003B50: CC404010 1C42F9CD
|
|
v_wmma_f32_16x16x16_f16 v[24:31], v[213:220], v[108:115], v[24:31]// 000000003B58: CC404018 1C62D9D5
|
|
v_wmma_f32_16x16x16_f16 v[32:39], v[213:220], v[116:123], v[32:39]// 000000003B60: CC404020 1C82E9D5
|
|
v_wmma_f32_16x16x16_f16 v[40:47], v[213:220], v[124:131], v[40:47]// 000000003B68: CC404028 1CA2F9D5
|
|
v_wmma_f32_16x16x16_f16 v[48:55], v[221:228], v[108:115], v[48:55]// 000000003B70: CC404030 1CC2D9DD
|
|
v_wmma_f32_16x16x16_f16 v[56:63], v[221:228], v[116:123], v[56:63]// 000000003B78: CC404038 1CE2E9DD
|
|
v_wmma_f32_16x16x16_f16 v[64:71], v[221:228], v[124:131], v[64:71]// 000000003B80: CC404040 1D02F9DD
|
|
v_and_b32_e32 v78, 0xf03fff, v78 // 000000003B88: 369C9CFF 00F03FFF
|
|
v_and_b32_e32 v79, 0xf03fff, v79 // 000000003B90: 369E9EFF 00F03FFF
|
|
s_and_b32 s12, 31, s27 // 000000003B98: 8B0C1B9F
|
|
s_and_b32 s66, s46, 0x8000 // 000000003B9C: 8B42FF2E 00008000
|
|
s_cmp_lg_u32 s6, s7 // 000000003BA4: BF070706
|
|
s_cmov_b32 s12, 0 // 000000003BA8: BE8C0280
|
|
s_cmp_eq_u32 s12, 0 // 000000003BAC: BF06800C
|
|
s_mov_b32 s13, 0 // 000000003BB0: BE8D0080
|
|
s_and_b32 s8, s46, 0x3fff // 000000003BB4: 8B08FF2E 00003FFF
|
|
s_load_b256 s[48:55], s[0:1], 0x58 // 000000003BBC: F40C0C00 F8000058
|
|
s_load_b32 s56, s[0:1], 0x78 // 000000003BC4: F4000E00 F8000078
|
|
v_mov_b32_e32 v75, s2 // 000000003BCC: 7E960202
|
|
v_mul_i32_i24_e32 v75, 0xffffffa0, v75 // 000000003BD0: 129696FF FFFFFFA0
|
|
v_add_co_u32 v75, vcc_lo, s24, v75 // 000000003BD8: D7006A4B 00029618
|
|
v_mov_b32_e32 v76, 0x60 // 000000003BE0: 7E9802FF 00000060
|
|
v_cmp_lt_u32_e64 s8, v75, v76 // 000000003BE8: D4490008 0002994B
|
|
v_cndmask_b32_e64 v75, v76, v75, s8 // 000000003BF0: D501004B 0022974C
|
|
v_lshrrev_b32_e32 v77, 5, v254 // 000000003BF8: 329BFC85
|
|
v_and_b32_e32 v77, 1, v77 // 000000003BFC: 369A9A81
|
|
v_lshrrev_b32_e32 v78, 4, v75 // 000000003C00: 329C9684
|
|
v_and_b32_e32 v78, 1, v78 // 000000003C04: 369C9C81
|
|
v_cmp_eq_u32_e64 s8, v78, v77 // 000000003C08: D44A0008 00029B4E
|
|
v_cndmask_b32_e64 v75, v76, v75, s8 // 000000003C10: D501004B 0022974C
|
|
v_lshrrev_b32_e32 v76, 4, v75 // 000000003C18: 32989684
|
|
v_lshlrev_b32_e32 v78, 0, v77 // 000000003C1C: 309C9A80
|
|
v_sub_nc_u32_e32 v76, v76, v78 // 000000003C20: 4C989D4C
|
|
v_lshrrev_b32_e32 v78, 3, v75 // 000000003C24: 329C9683
|
|
v_lshrrev_b32_e32 v79, 0, v254 // 000000003C28: 329FFC80
|
|
v_and_b32_e32 v79, 15, v79 // 000000003C2C: 369E9E8F
|
|
v_lshrrev_b32_e32 v79, 3, v79 // 000000003C30: 329E9E83
|
|
v_lshlrev_b32_e32 v77, 1, v77 // 000000003C34: 309A9A81
|
|
v_add_co_u32 v79, vcc_lo, v77, v79 // 000000003C38: D7006A4F 00029F4D
|
|
v_sub_nc_u32_e32 v78, v78, v79 // 000000003C40: 4C9C9F4E
|
|
v_and_b32_e32 v77, 0, v75 // 000000003C44: 369A9680
|
|
v_lshrrev_b32_e32 v77, 3, v77 // 000000003C48: 329A9A83
|
|
v_and_b32_e32 v79, 7, v75 // 000000003C4C: 369E9687
|
|
v_cmp_eq_u32_e64 vcc_lo, v79, 1 // 000000003C50: D44A006A 0001034F
|
|
v_lshrrev_b32_e32 v76, 5, v254 // 000000003C58: 3299FC85
|
|
v_lshrrev_b32_e32 v77, 1, v76 // 000000003C5C: 329A9881
|
|
v_mul_lo_u32 v77, 16, v77 // 000000003C60: D72C004D 00029A90
|
|
v_and_b32_e32 v73, 31, v254 // 000000003C68: 3693FC9F
|
|
v_lshrrev_b32_e32 v73, 4, v73 // 000000003C6C: 32929284
|
|
v_add_lshl_u32 v73, v77, v73, 0 // 000000003C70: D6470049 0202934D
|
|
v_mul_lo_u32 v74, v73, s38 // 000000003C78: D72C004A 00004D49
|
|
v_mul_lo_u32 v75, v73, s36 // 000000003C80: D72C004B 00004949
|
|
v_and_b32_e32 v72, 1, v76 // 000000003C88: 36909881
|
|
v_mul_lo_u32 v72, 16, v72 // 000000003C8C: D72C0048 00029090
|
|
v_and_b32_e32 v77, 15, v254 // 000000003C94: 369BFC8F
|
|
v_add_lshl_u32 v72, v77, v72, 0 // 000000003C98: D6470048 0202914D
|
|
s_mul_i32 s8, 0x60, s2 // 000000003CA0: 960802FF 00000060
|
|
v_add_nc_u32_e32 v72, s8, v72 // 000000003CA8: 4A909008
|
|
s_mul_i32 s8, 0x60, s3 // 000000003CAC: 960803FF 00000060
|
|
v_add_nc_u32_e32 v73, s8, v73 // 000000003CB4: 4A929208
|
|
s_waitcnt lgkmcnt(0) // 000000003CB8: BF89FC07
|
|
s_and_b32 s8, s46, 0x3fff // 000000003CBC: 8B08FF2E 00003FFF
|
|
s_mov_b32 s33, 0 // 000000003CC4: BEA10080
|
|
s_mul_i32 s32, 0x555, s24 // 000000003CC8: 962018FF 00000555
|
|
s_lshl_b64 s[32:33], s[32:33], 16 // 000000003CD0: 84A09020
|
|
s_mul_i32 s31, s24, 0x5556 // 000000003CD4: 961FFF18 00005556
|
|
s_add_u32 s32, s31, s32 // 000000003CDC: 8020201F
|
|
s_addc_u32 s33, s33, 0 // 000000003CE0: 82218021
|
|
s_lshr_b64 s[32:33], s[32:33], 33 // 000000003CE4: 85A0A120
|
|
s_mov_b32 s31, s32 // 000000003CE8: BE9F0020
|
|
s_mul_i32 s32, s31, 0x60 // 000000003CEC: 9620FF1F 00000060
|
|
s_sub_u32 s30, s24, s32 // 000000003CF4: 809E2018
|
|
s_add_u32 s31, -1, s14 // 000000003CF8: 801F0EC1
|
|
s_cmp_ge_u32 s2, s31 // 000000003CFC: BF091F02
|
|
s_cselect_b32 s30, s30, 0 // 000000003D00: 981E801E
|
|
s_cmpk_gt_u32 s30, 0x0 // 000000003D04: B59E0000
|
|
s_mov_b32 s33, 0 // 000000003D08: BEA10080
|
|
s_mul_i32 s32, 0x555, s25 // 000000003D0C: 962019FF 00000555
|
|
s_lshl_b64 s[32:33], s[32:33], 16 // 000000003D14: 84A09020
|
|
s_mul_i32 s31, s25, 0x5556 // 000000003D18: 961FFF19 00005556
|
|
s_add_u32 s32, s31, s32 // 000000003D20: 8020201F
|
|
s_addc_u32 s33, s33, 0 // 000000003D24: 82218021
|
|
s_lshr_b64 s[32:33], s[32:33], 33 // 000000003D28: 85A0A120
|
|
s_mov_b32 s31, s32 // 000000003D2C: BE9F0020
|
|
s_mul_i32 s32, s31, 0x60 // 000000003D30: 9620FF1F 00000060
|
|
s_sub_u32 s30, s25, s32 // 000000003D38: 809E2019
|
|
s_add_u32 s31, -1, s15 // 000000003D3C: 801F0FC1
|
|
s_cmp_ge_u32 s3, s31 // 000000003D40: BF091F03
|
|
s_cselect_b32 s30, s30, 0 // 000000003D44: 981E801E
|
|
s_cmpk_gt_u32 s30, 0x0 // 000000003D48: B59E0000
|
|
s_mov_b64 s[32:33], s[48:49] // 000000003D4C: BEA00130
|
|
s_mov_b32 s35, 0x31004000 // 000000003D50: BEA300FF 31004000
|
|
s_mov_b32 s34, 0 // 000000003D58: BEA20080
|
|
s_mul_i32 s34, 4, s34 // 000000003D5C: 96222284
|
|
s_add_u32 s8, s4, 1 // 000000003D60: 80088104
|
|
s_mul_i32 s8, s53, s8 // 000000003D64: 96080835
|
|
s_cmp_eq_u32 s8, 0 // 000000003D68: BF068008
|
|
s_cselect_b32 s8, s24, s8 // 000000003D6C: 98080818
|
|
s_mov_b64 s[40:41], s[50:51] // 000000003D70: BEA80132
|
|
s_mov_b32 s43, 0x31004000 // 000000003D74: BEAB00FF 31004000
|
|
s_mov_b32 s42, 0 // 000000003D7C: BEAA0080
|
|
s_mul_i32 s8, 0x60, s2 // 000000003D80: 960802FF 00000060
|
|
v_add_nc_u32_e32 v80, s8, v254 // 000000003D88: 4AA1FC08
|
|
s_mul_i32 s42, 4, s42 // 000000003D8C: 962A2A84
|
|
s_mul_i32 s8, s53, s4 // 000000003D90: 96080435
|
|
v_add_nc_u32_e32 v78, s8, v80 // 000000003D94: 4A9CA008
|
|
v_lshlrev_b32_e32 v78, 2, v78 // 000000003D98: 309C9C82
|
|
v_lshlrev_b32_e32 v79, 2, v80 // 000000003D9C: 309EA082
|
|
s_mul_i32 s8, 0x60, s3 // 000000003DA0: 960803FF 00000060
|
|
v_add_nc_u32_e32 v80, s8, v254 // 000000003DA8: 4AA1FC08
|
|
buffer_load_b32 v76, v78, s[40:43], 0 offen // 000000003DAC: E0500000 804A4C4E
|
|
buffer_load_b32 v77, v79, s[32:35], 0 offen // 000000003DB4: E0500000 80484D4F
|
|
v_lshlrev_b32_e32 v80, 2, v254 // 000000003DBC: 30A1FC82
|
|
s_barrier // 000000003DC0: BFBD0000
|
|
s_waitcnt vmcnt(1) // 000000003DC4: BF8907F7
|
|
ds_store_b32 v80, v76 // 000000003DC8: D8340000 00004C50
|
|
v_cmp_gt_u32_e64 s48, s34, 0 // 000000003DD0: D44C0030 00010022
|
|
s_waitcnt vmcnt(0) // 000000003DD8: BF8903F7
|
|
v_cndmask_b32_e64 v77, 1.0, v77, s48 // 000000003DDC: D501004D 00C29AF2
|
|
ds_store_b32 v80, v77 offset:512 // 000000003DE4: D8340200 00004D50
|
|
s_add_u32 s12, s12, 0x6254 // 000000003DEC: 800CFF0C 00006254
|
|
s_addc_u32 s13, s13, 0 // 000000003DF4: 820D800D
|
|
s_mov_b32 s35, 0 // 000000003DF8: BEA30080
|
|
s_mul_i32 s34, 0x555, s24 // 000000003DFC: 962218FF 00000555
|
|
s_lshl_b64 s[34:35], s[34:35], 16 // 000000003E04: 84A29022
|
|
s_mul_i32 s33, s24, 0x5556 // 000000003E08: 9621FF18 00005556
|
|
s_add_u32 s34, s33, s34 // 000000003E10: 80222221
|
|
s_addc_u32 s35, s35, 0 // 000000003E14: 82238023
|
|
s_lshr_b64 s[34:35], s[34:35], 33 // 000000003E18: 85A2A122
|
|
s_mov_b32 s33, s34 // 000000003E1C: BEA10022
|
|
s_mul_i32 s34, s33, 0x60 // 000000003E20: 9622FF21 00000060
|
|
s_sub_u32 s32, s24, s34 // 000000003E28: 80A02218
|
|
s_add_u32 s33, -1, s14 // 000000003E2C: 80210EC1
|
|
s_cmp_ge_u32 s2, s33 // 000000003E30: BF092102
|
|
s_cselect_b32 s32, s32, 0 // 000000003E34: 98208020
|
|
s_cmpk_gt_u32 s32, 0x0 // 000000003E38: B5A00000
|
|
v_mov_b32_e32 v78, 0x80000000 // 000000003E3C: 7E9C02FF 80000000
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000003E44: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000003E4C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000003E54: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000003E58: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v116, v72, s32 // 000000003E60: D5260074 00004148
|
|
v_lshlrev_b32_e32 v116, 2, v116 // 000000003E68: 30E8E882
|
|
s_waitcnt lgkmcnt(0) // 000000003E6C: BF89FC07
|
|
s_barrier // 000000003E70: BFBD0000
|
|
ds_load_b32 v113, v116 // 000000003E74: D8D80000 71000074
|
|
ds_load_b32 v114, v116 offset:512 // 000000003E7C: D8D80200 72000074
|
|
v_add_lshl_u32 v115, v75, v72, 1 // 000000003E84: D6470073 0206914B
|
|
v_cndmask_b32_e64 v115, v78, v115, s34 // 000000003E8C: D5010073 008AE74E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000003E94: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000003E9C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000003EA4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000003EAC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000003EB0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v120, v76, s32 // 000000003EB8: D5260078 0000414C
|
|
v_lshlrev_b32_e32 v120, 2, v120 // 000000003EC0: 30F0F082
|
|
ds_load_b32 v117, v120 // 000000003EC4: D8D80000 75000078
|
|
ds_load_b32 v118, v120 offset:512 // 000000003ECC: D8D80200 76000078
|
|
v_add_lshl_u32 v119, v75, v76, 1 // 000000003ED4: D6470077 0206994B
|
|
v_cndmask_b32_e64 v119, v78, v119, s34 // 000000003EDC: D5010077 008AEF4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000003EE4: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000003EEC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000003EF4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000003EFC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000003F00: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v124, v76, s32 // 000000003F08: D526007C 0000414C
|
|
v_lshlrev_b32_e32 v124, 2, v124 // 000000003F10: 30F8F882
|
|
ds_load_b32 v121, v124 // 000000003F14: D8D80000 7900007C
|
|
ds_load_b32 v122, v124 offset:512 // 000000003F1C: D8D80200 7A00007C
|
|
v_add_lshl_u32 v123, v75, v76, 1 // 000000003F24: D647007B 0206994B
|
|
v_cndmask_b32_e64 v123, v78, v123, s34 // 000000003F2C: D501007B 008AF74E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000003F34: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 000000003F3C: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000003F40: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000003F48: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 000000003F4C: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000003F54: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000003F5C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000003F64: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000003F68: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v126, v72, s32 // 000000003F70: D526007E 00004148
|
|
v_lshlrev_b32_e32 v126, 2, v126 // 000000003F78: 30FCFC82
|
|
v_add_lshl_u32 v125, v75, v72, 1 // 000000003F7C: D647007D 0206914B
|
|
v_cndmask_b32_e64 v125, v78, v125, s34 // 000000003F84: D501007D 008AFB4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000003F8C: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000003F94: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000003F9C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000003FA4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000003FA8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v128, v76, s32 // 000000003FB0: D5260080 0000414C
|
|
v_lshlrev_b32_e32 v128, 2, v128 // 000000003FB8: 31010082
|
|
v_add_lshl_u32 v127, v75, v76, 1 // 000000003FBC: D647007F 0206994B
|
|
v_cndmask_b32_e64 v127, v78, v127, s34 // 000000003FC4: D501007F 008AFF4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000003FCC: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000003FD4: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000003FDC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000003FE4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000003FE8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v130, v76, s32 // 000000003FF0: D5260082 0000414C
|
|
v_lshlrev_b32_e32 v130, 2, v130 // 000000003FF8: 31050482
|
|
v_add_lshl_u32 v129, v75, v76, 1 // 000000003FFC: D6470081 0206994B
|
|
v_cndmask_b32_e64 v129, v78, v129, s34 // 000000004004: D5010081 008B034E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 00000000400C: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 000000004014: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004018: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000004020: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 000000004024: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 00000000402C: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004034: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000403C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004040: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v132, v72, s32 // 000000004048: D5260084 00004148
|
|
v_lshlrev_b32_e32 v132, 2, v132 // 000000004050: 31090882
|
|
v_add_lshl_u32 v131, v75, v72, 1 // 000000004054: D6470083 0206914B
|
|
v_cndmask_b32_e64 v131, v78, v131, s34 // 00000000405C: D5010083 008B074E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000004064: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000406C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004074: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000407C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004080: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v134, v76, s32 // 000000004088: D5260086 0000414C
|
|
v_lshlrev_b32_e32 v134, 2, v134 // 000000004090: 310D0C82
|
|
v_add_lshl_u32 v133, v75, v76, 1 // 000000004094: D6470085 0206994B
|
|
v_cndmask_b32_e64 v133, v78, v133, s34 // 00000000409C: D5010085 008B0B4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 0000000040A4: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000040AC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000040B4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000040BC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000040C0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v136, v76, s32 // 0000000040C8: D5260088 0000414C
|
|
v_lshlrev_b32_e32 v136, 2, v136 // 0000000040D0: 31111082
|
|
v_add_lshl_u32 v135, v75, v76, 1 // 0000000040D4: D6470087 0206994B
|
|
v_cndmask_b32_e64 v135, v78, v135, s34 // 0000000040DC: D5010087 008B0F4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 0000000040E4: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 0000000040EC: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 0000000040F0: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 0000000040F8: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 0000000040FC: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000004104: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000410C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004114: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004118: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v138, v72, s32 // 000000004120: D526008A 00004148
|
|
v_lshlrev_b32_e32 v138, 2, v138 // 000000004128: 31151482
|
|
v_add_lshl_u32 v137, v75, v72, 1 // 00000000412C: D6470089 0206914B
|
|
v_cndmask_b32_e64 v137, v78, v137, s34 // 000000004134: D5010089 008B134E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 00000000413C: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004144: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000414C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004154: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004158: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v140, v76, s32 // 000000004160: D526008C 0000414C
|
|
v_lshlrev_b32_e32 v140, 2, v140 // 000000004168: 31191882
|
|
v_add_lshl_u32 v139, v75, v76, 1 // 00000000416C: D647008B 0206994B
|
|
v_cndmask_b32_e64 v139, v78, v139, s34 // 000000004174: D501008B 008B174E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 00000000417C: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004184: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000418C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004194: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004198: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v142, v76, s32 // 0000000041A0: D526008E 0000414C
|
|
v_lshlrev_b32_e32 v142, 2, v142 // 0000000041A8: 311D1C82
|
|
v_add_lshl_u32 v141, v75, v76, 1 // 0000000041AC: D647008D 0206994B
|
|
v_cndmask_b32_e64 v141, v78, v141, s34 // 0000000041B4: D501008D 008B1B4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 0000000041BC: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 0000000041C4: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 0000000041C8: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 0000000041D0: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 0000000041D4: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 0000000041DC: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000041E4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000041EC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000041F0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v144, v72, s32 // 0000000041F8: D5260090 00004148
|
|
v_lshlrev_b32_e32 v144, 2, v144 // 000000004200: 31212082
|
|
v_add_lshl_u32 v143, v75, v72, 1 // 000000004204: D647008F 0206914B
|
|
v_cndmask_b32_e64 v143, v78, v143, s34 // 00000000420C: D501008F 008B1F4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000004214: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000421C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004224: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000422C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004230: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v146, v76, s32 // 000000004238: D5260092 0000414C
|
|
v_lshlrev_b32_e32 v146, 2, v146 // 000000004240: 31252482
|
|
v_add_lshl_u32 v145, v75, v76, 1 // 000000004244: D6470091 0206994B
|
|
v_cndmask_b32_e64 v145, v78, v145, s34 // 00000000424C: D5010091 008B234E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000004254: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000425C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004264: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000426C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004270: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v148, v76, s32 // 000000004278: D5260094 0000414C
|
|
v_lshlrev_b32_e32 v148, 2, v148 // 000000004280: 31292882
|
|
v_add_lshl_u32 v147, v75, v76, 1 // 000000004284: D6470093 0206994B
|
|
v_cndmask_b32_e64 v147, v78, v147, s34 // 00000000428C: D5010093 008B274E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000004294: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 00000000429C: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 0000000042A0: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 0000000042A8: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 0000000042AC: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 0000000042B4: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000042BC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000042C4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000042C8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v150, v72, s32 // 0000000042D0: D5260096 00004148
|
|
v_lshlrev_b32_e32 v150, 2, v150 // 0000000042D8: 312D2C82
|
|
v_add_lshl_u32 v149, v75, v72, 1 // 0000000042DC: D6470095 0206914B
|
|
v_cndmask_b32_e64 v149, v78, v149, s34 // 0000000042E4: D5010095 008B2B4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 0000000042EC: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000042F4: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000042FC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004304: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004308: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v152, v76, s32 // 000000004310: D5260098 0000414C
|
|
v_lshlrev_b32_e32 v152, 2, v152 // 000000004318: 31313082
|
|
v_add_lshl_u32 v151, v75, v76, 1 // 00000000431C: D6470097 0206994B
|
|
v_cndmask_b32_e64 v151, v78, v151, s34 // 000000004324: D5010097 008B2F4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 00000000432C: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004334: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000433C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004344: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004348: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v154, v76, s32 // 000000004350: D526009A 0000414C
|
|
v_lshlrev_b32_e32 v154, 2, v154 // 000000004358: 31353482
|
|
v_add_lshl_u32 v153, v75, v76, 1 // 00000000435C: D6470099 0206994B
|
|
v_cndmask_b32_e64 v153, v78, v153, s34 // 000000004364: D5010099 008B334E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 00000000436C: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 000000004374: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004378: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000004380: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 000000004384: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 00000000438C: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004394: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000439C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000043A0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v156, v72, s32 // 0000000043A8: D526009C 00004148
|
|
v_lshlrev_b32_e32 v156, 2, v156 // 0000000043B0: 31393882
|
|
v_add_lshl_u32 v155, v75, v72, 1 // 0000000043B4: D647009B 0206914B
|
|
v_cndmask_b32_e64 v155, v78, v155, s34 // 0000000043BC: D501009B 008B374E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 0000000043C4: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000043CC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000043D4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000043DC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000043E0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v158, v76, s32 // 0000000043E8: D526009E 0000414C
|
|
v_lshlrev_b32_e32 v158, 2, v158 // 0000000043F0: 313D3C82
|
|
v_add_lshl_u32 v157, v75, v76, 1 // 0000000043F4: D647009D 0206994B
|
|
v_cndmask_b32_e64 v157, v78, v157, s34 // 0000000043FC: D501009D 008B3B4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000004404: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000440C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004414: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000441C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004420: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v160, v76, s32 // 000000004428: D52600A0 0000414C
|
|
v_lshlrev_b32_e32 v160, 2, v160 // 000000004430: 31414082
|
|
v_add_lshl_u32 v159, v75, v76, 1 // 000000004434: D647009F 0206994B
|
|
v_cndmask_b32_e64 v159, v78, v159, s34 // 00000000443C: D501009F 008B3F4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000004444: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 00000000444C: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004450: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000004458: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 00000000445C: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000004464: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000446C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004474: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004478: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v162, v72, s32 // 000000004480: D52600A2 00004148
|
|
v_lshlrev_b32_e32 v162, 2, v162 // 000000004488: 31454482
|
|
v_add_lshl_u32 v161, v75, v72, 1 // 00000000448C: D64700A1 0206914B
|
|
v_cndmask_b32_e64 v161, v78, v161, s34 // 000000004494: D50100A1 008B434E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 00000000449C: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000044A4: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000044AC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000044B4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000044B8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v164, v76, s32 // 0000000044C0: D52600A4 0000414C
|
|
v_lshlrev_b32_e32 v164, 2, v164 // 0000000044C8: 31494882
|
|
v_add_lshl_u32 v163, v75, v76, 1 // 0000000044CC: D64700A3 0206994B
|
|
v_cndmask_b32_e64 v163, v78, v163, s34 // 0000000044D4: D50100A3 008B474E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 0000000044DC: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000044E4: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000044EC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000044F4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000044F8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v166, v76, s32 // 000000004500: D52600A6 0000414C
|
|
v_lshlrev_b32_e32 v166, 2, v166 // 000000004508: 314D4C82
|
|
v_add_lshl_u32 v165, v75, v76, 1 // 00000000450C: D64700A5 0206994B
|
|
v_cndmask_b32_e64 v165, v78, v165, s34 // 000000004514: D50100A5 008B4B4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 18 // 00000000451C: D7006A49 00012549
|
|
s_mul_i32 s32, s38, 18 // 000000004524: 96209226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004528: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 18 // 000000004530: 96209224
|
|
v_add_nc_i32 v75, v75, s32 // 000000004534: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 00000000453C: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004544: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000454C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004550: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v168, v72, s32 // 000000004558: D52600A8 00004148
|
|
v_lshlrev_b32_e32 v168, 2, v168 // 000000004560: 31515082
|
|
v_add_lshl_u32 v167, v75, v72, 1 // 000000004564: D64700A7 0206914B
|
|
v_cndmask_b32_e64 v167, v78, v167, s34 // 00000000456C: D50100A7 008B4F4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000004574: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000457C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004584: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000458C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004590: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v170, v76, s32 // 000000004598: D52600AA 0000414C
|
|
v_lshlrev_b32_e32 v170, 2, v170 // 0000000045A0: 31555482
|
|
v_add_lshl_u32 v169, v75, v76, 1 // 0000000045A4: D64700A9 0206994B
|
|
v_cndmask_b32_e64 v169, v78, v169, s34 // 0000000045AC: D50100A9 008B534E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 0000000045B4: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000045BC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000045C4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000045CC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000045D0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v172, v76, s32 // 0000000045D8: D52600AC 0000414C
|
|
v_lshlrev_b32_e32 v172, 2, v172 // 0000000045E0: 31595882
|
|
v_add_lshl_u32 v171, v75, v76, 1 // 0000000045E4: D64700AB 0206994B
|
|
v_cndmask_b32_e64 v171, v78, v171, s34 // 0000000045EC: D50100AB 008B574E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 0000000045F4: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 0000000045FC: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004600: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000004608: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 00000000460C: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000004614: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000461C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004624: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004628: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v174, v72, s32 // 000000004630: D52600AE 00004148
|
|
v_lshlrev_b32_e32 v174, 2, v174 // 000000004638: 315D5C82
|
|
v_add_lshl_u32 v173, v75, v72, 1 // 00000000463C: D64700AD 0206914B
|
|
v_cndmask_b32_e64 v173, v78, v173, s34 // 000000004644: D50100AD 008B5B4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 00000000464C: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004654: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000465C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004664: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004668: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v176, v76, s32 // 000000004670: D52600B0 0000414C
|
|
v_lshlrev_b32_e32 v176, 2, v176 // 000000004678: 31616082
|
|
v_add_lshl_u32 v175, v75, v76, 1 // 00000000467C: D64700AF 0206994B
|
|
v_cndmask_b32_e64 v175, v78, v175, s34 // 000000004684: D50100AF 008B5F4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 00000000468C: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004694: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000469C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000046A4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000046A8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v178, v76, s32 // 0000000046B0: D52600B2 0000414C
|
|
v_lshlrev_b32_e32 v178, 2, v178 // 0000000046B8: 31656482
|
|
v_add_lshl_u32 v177, v75, v76, 1 // 0000000046BC: D64700B1 0206994B
|
|
v_cndmask_b32_e64 v177, v78, v177, s34 // 0000000046C4: D50100B1 008B634E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 0000000046CC: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 0000000046D4: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 0000000046D8: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 0000000046E0: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 0000000046E4: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 0000000046EC: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000046F4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000046FC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004700: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v180, v72, s32 // 000000004708: D52600B4 00004148
|
|
v_lshlrev_b32_e32 v180, 2, v180 // 000000004710: 31696882
|
|
v_add_lshl_u32 v179, v75, v72, 1 // 000000004714: D64700B3 0206914B
|
|
v_cndmask_b32_e64 v179, v78, v179, s34 // 00000000471C: D50100B3 008B674E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000004724: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000472C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004734: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000473C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004740: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v182, v76, s32 // 000000004748: D52600B6 0000414C
|
|
v_lshlrev_b32_e32 v182, 2, v182 // 000000004750: 316D6C82
|
|
v_add_lshl_u32 v181, v75, v76, 1 // 000000004754: D64700B5 0206994B
|
|
v_cndmask_b32_e64 v181, v78, v181, s34 // 00000000475C: D50100B5 008B6B4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000004764: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000476C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004774: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000477C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004780: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v184, v76, s32 // 000000004788: D52600B8 0000414C
|
|
v_lshlrev_b32_e32 v184, 2, v184 // 000000004790: 31717082
|
|
v_add_lshl_u32 v183, v75, v76, 1 // 000000004794: D64700B7 0206994B
|
|
v_cndmask_b32_e64 v183, v78, v183, s34 // 00000000479C: D50100B7 008B6F4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 0000000047A4: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 0000000047AC: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 0000000047B0: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 0000000047B8: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 0000000047BC: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 0000000047C4: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000047CC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000047D4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000047D8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v186, v72, s32 // 0000000047E0: D52600BA 00004148
|
|
v_lshlrev_b32_e32 v186, 2, v186 // 0000000047E8: 31757482
|
|
v_add_lshl_u32 v185, v75, v72, 1 // 0000000047EC: D64700B9 0206914B
|
|
v_cndmask_b32_e64 v185, v78, v185, s34 // 0000000047F4: D50100B9 008B734E
|
|
v_mov_b32_e32 v79, v0 // 0000000047FC: 7E9E0300
|
|
v_mov_b32_e32 v80, v8 // 000000004800: 7EA00308
|
|
v_mov_b32_e32 v81, v16 // 000000004804: 7EA20310
|
|
v_mov_b32_e32 v82, v1 // 000000004808: 7EA40301
|
|
v_mov_b32_e32 v83, v9 // 00000000480C: 7EA60309
|
|
v_mov_b32_e32 v84, v17 // 000000004810: 7EA80311
|
|
v_mov_b32_e32 v85, v2 // 000000004814: 7EAA0302
|
|
v_mov_b32_e32 v86, v10 // 000000004818: 7EAC030A
|
|
v_mov_b32_e32 v87, v18 // 00000000481C: 7EAE0312
|
|
v_mov_b32_e32 v88, v3 // 000000004820: 7EB00303
|
|
v_mov_b32_e32 v89, v11 // 000000004824: 7EB2030B
|
|
v_mov_b32_e32 v90, v19 // 000000004828: 7EB40313
|
|
v_mov_b32_e32 v91, v4 // 00000000482C: 7EB60304
|
|
v_mov_b32_e32 v92, v12 // 000000004830: 7EB8030C
|
|
v_mov_b32_e32 v93, v20 // 000000004834: 7EBA0314
|
|
v_mov_b32_e32 v94, v5 // 000000004838: 7EBC0305
|
|
v_mov_b32_e32 v95, v13 // 00000000483C: 7EBE030D
|
|
v_mov_b32_e32 v96, v21 // 000000004840: 7EC00315
|
|
v_mov_b32_e32 v97, v6 // 000000004844: 7EC20306
|
|
v_mov_b32_e32 v98, v14 // 000000004848: 7EC4030E
|
|
v_mov_b32_e32 v99, v22 // 00000000484C: 7EC60316
|
|
v_mov_b32_e32 v100, v7 // 000000004850: 7EC80307
|
|
v_mov_b32_e32 v101, v15 // 000000004854: 7ECA030F
|
|
v_mov_b32_e32 v102, v23 // 000000004858: 7ECC0317
|
|
v_mov_b32_e32 v103, v24 // 00000000485C: 7ECE0318
|
|
v_mov_b32_e32 v104, v32 // 000000004860: 7ED00320
|
|
v_mov_b32_e32 v105, v40 // 000000004864: 7ED20328
|
|
v_mov_b32_e32 v106, v25 // 000000004868: 7ED40319
|
|
v_mov_b32_e32 v107, v33 // 00000000486C: 7ED60321
|
|
v_mov_b32_e32 v108, v41 // 000000004870: 7ED80329
|
|
v_mov_b32_e32 v109, v26 // 000000004874: 7EDA031A
|
|
v_mov_b32_e32 v110, v34 // 000000004878: 7EDC0322
|
|
v_mov_b32_e32 v111, v42 // 00000000487C: 7EDE032A
|
|
v_mov_b32_e32 v112, v27 // 000000004880: 7EE0031B
|
|
s_waitcnt lgkmcnt(0) // 000000004884: BF89FC07
|
|
v_mul_f32_e32 v79, v114, v79 // 000000004888: 109E9F72
|
|
v_add_f32_e32 v76, v113, v79 // 00000000488C: 06989F71
|
|
v_mov_b32_e32 v79, v76 // 000000004890: 7E9E034C
|
|
v_cvt_f16_f32_e32 v79, v79 // 000000004894: 7E9E154F
|
|
buffer_store_b16 v79, v115, s[16:19], 0 offen // 000000004898: E0640000 80444F73
|
|
v_mul_f32_e32 v80, v118, v80 // 0000000048A0: 10A0A176
|
|
v_add_f32_e32 v76, v117, v80 // 0000000048A4: 0698A175
|
|
v_mov_b32_e32 v80, v76 // 0000000048A8: 7EA0034C
|
|
v_cvt_f16_f32_e32 v80, v80 // 0000000048AC: 7EA01550
|
|
buffer_store_b16 v80, v119, s[16:19], 0 offen // 0000000048B0: E0640000 80445077
|
|
v_mul_f32_e32 v81, v122, v81 // 0000000048B8: 10A2A37A
|
|
v_add_f32_e32 v76, v121, v81 // 0000000048BC: 0698A379
|
|
v_mov_b32_e32 v81, v76 // 0000000048C0: 7EA2034C
|
|
v_cvt_f16_f32_e32 v81, v81 // 0000000048C4: 7EA21551
|
|
buffer_store_b16 v81, v123, s[16:19], 0 offen // 0000000048C8: E0640000 8044517B
|
|
v_mul_f32_e32 v82, v114, v82 // 0000000048D0: 10A4A572
|
|
v_add_f32_e32 v76, v113, v82 // 0000000048D4: 0698A571
|
|
v_mov_b32_e32 v82, v76 // 0000000048D8: 7EA4034C
|
|
v_cvt_f16_f32_e32 v82, v82 // 0000000048DC: 7EA41552
|
|
buffer_store_b16 v82, v125, s[16:19], 0 offen // 0000000048E0: E0640000 8044527D
|
|
v_mul_f32_e32 v83, v118, v83 // 0000000048E8: 10A6A776
|
|
v_add_f32_e32 v76, v117, v83 // 0000000048EC: 0698A775
|
|
v_mov_b32_e32 v83, v76 // 0000000048F0: 7EA6034C
|
|
v_cvt_f16_f32_e32 v83, v83 // 0000000048F4: 7EA61553
|
|
buffer_store_b16 v83, v127, s[16:19], 0 offen // 0000000048F8: E0640000 8044537F
|
|
v_mul_f32_e32 v84, v122, v84 // 000000004900: 10A8A97A
|
|
v_add_f32_e32 v76, v121, v84 // 000000004904: 0698A979
|
|
v_mov_b32_e32 v84, v76 // 000000004908: 7EA8034C
|
|
v_cvt_f16_f32_e32 v84, v84 // 00000000490C: 7EA81554
|
|
buffer_store_b16 v84, v129, s[16:19], 0 offen // 000000004910: E0640000 80445481
|
|
v_mul_f32_e32 v85, v114, v85 // 000000004918: 10AAAB72
|
|
v_add_f32_e32 v76, v113, v85 // 00000000491C: 0698AB71
|
|
v_mov_b32_e32 v85, v76 // 000000004920: 7EAA034C
|
|
v_cvt_f16_f32_e32 v85, v85 // 000000004924: 7EAA1555
|
|
buffer_store_b16 v85, v131, s[16:19], 0 offen // 000000004928: E0640000 80445583
|
|
v_mul_f32_e32 v86, v118, v86 // 000000004930: 10ACAD76
|
|
v_add_f32_e32 v76, v117, v86 // 000000004934: 0698AD75
|
|
v_mov_b32_e32 v86, v76 // 000000004938: 7EAC034C
|
|
v_cvt_f16_f32_e32 v86, v86 // 00000000493C: 7EAC1556
|
|
buffer_store_b16 v86, v133, s[16:19], 0 offen // 000000004940: E0640000 80445685
|
|
v_mul_f32_e32 v87, v122, v87 // 000000004948: 10AEAF7A
|
|
v_add_f32_e32 v76, v121, v87 // 00000000494C: 0698AF79
|
|
v_mov_b32_e32 v87, v76 // 000000004950: 7EAE034C
|
|
v_cvt_f16_f32_e32 v87, v87 // 000000004954: 7EAE1557
|
|
buffer_store_b16 v87, v135, s[16:19], 0 offen // 000000004958: E0640000 80445787
|
|
v_mul_f32_e32 v88, v114, v88 // 000000004960: 10B0B172
|
|
v_add_f32_e32 v76, v113, v88 // 000000004964: 0698B171
|
|
v_mov_b32_e32 v88, v76 // 000000004968: 7EB0034C
|
|
v_cvt_f16_f32_e32 v88, v88 // 00000000496C: 7EB01558
|
|
buffer_store_b16 v88, v137, s[16:19], 0 offen // 000000004970: E0640000 80445889
|
|
v_mul_f32_e32 v89, v118, v89 // 000000004978: 10B2B376
|
|
v_add_f32_e32 v76, v117, v89 // 00000000497C: 0698B375
|
|
v_mov_b32_e32 v89, v76 // 000000004980: 7EB2034C
|
|
v_cvt_f16_f32_e32 v89, v89 // 000000004984: 7EB21559
|
|
buffer_store_b16 v89, v139, s[16:19], 0 offen // 000000004988: E0640000 8044598B
|
|
v_mul_f32_e32 v90, v122, v90 // 000000004990: 10B4B57A
|
|
v_add_f32_e32 v76, v121, v90 // 000000004994: 0698B579
|
|
v_mov_b32_e32 v90, v76 // 000000004998: 7EB4034C
|
|
v_cvt_f16_f32_e32 v90, v90 // 00000000499C: 7EB4155A
|
|
buffer_store_b16 v90, v141, s[16:19], 0 offen // 0000000049A0: E0640000 80445A8D
|
|
v_mul_f32_e32 v91, v114, v91 // 0000000049A8: 10B6B772
|
|
v_add_f32_e32 v76, v113, v91 // 0000000049AC: 0698B771
|
|
v_mov_b32_e32 v91, v76 // 0000000049B0: 7EB6034C
|
|
v_cvt_f16_f32_e32 v91, v91 // 0000000049B4: 7EB6155B
|
|
buffer_store_b16 v91, v143, s[16:19], 0 offen // 0000000049B8: E0640000 80445B8F
|
|
v_mul_f32_e32 v92, v118, v92 // 0000000049C0: 10B8B976
|
|
v_add_f32_e32 v76, v117, v92 // 0000000049C4: 0698B975
|
|
v_mov_b32_e32 v92, v76 // 0000000049C8: 7EB8034C
|
|
v_cvt_f16_f32_e32 v92, v92 // 0000000049CC: 7EB8155C
|
|
buffer_store_b16 v92, v145, s[16:19], 0 offen // 0000000049D0: E0640000 80445C91
|
|
v_mul_f32_e32 v93, v122, v93 // 0000000049D8: 10BABB7A
|
|
v_add_f32_e32 v76, v121, v93 // 0000000049DC: 0698BB79
|
|
v_mov_b32_e32 v93, v76 // 0000000049E0: 7EBA034C
|
|
v_cvt_f16_f32_e32 v93, v93 // 0000000049E4: 7EBA155D
|
|
buffer_store_b16 v93, v147, s[16:19], 0 offen // 0000000049E8: E0640000 80445D93
|
|
v_mul_f32_e32 v94, v114, v94 // 0000000049F0: 10BCBD72
|
|
v_add_f32_e32 v76, v113, v94 // 0000000049F4: 0698BD71
|
|
v_mov_b32_e32 v94, v76 // 0000000049F8: 7EBC034C
|
|
v_cvt_f16_f32_e32 v94, v94 // 0000000049FC: 7EBC155E
|
|
buffer_store_b16 v94, v149, s[16:19], 0 offen // 000000004A00: E0640000 80445E95
|
|
v_mul_f32_e32 v95, v118, v95 // 000000004A08: 10BEBF76
|
|
v_add_f32_e32 v76, v117, v95 // 000000004A0C: 0698BF75
|
|
v_mov_b32_e32 v95, v76 // 000000004A10: 7EBE034C
|
|
v_cvt_f16_f32_e32 v95, v95 // 000000004A14: 7EBE155F
|
|
buffer_store_b16 v95, v151, s[16:19], 0 offen // 000000004A18: E0640000 80445F97
|
|
v_mul_f32_e32 v96, v122, v96 // 000000004A20: 10C0C17A
|
|
v_add_f32_e32 v76, v121, v96 // 000000004A24: 0698C179
|
|
v_mov_b32_e32 v96, v76 // 000000004A28: 7EC0034C
|
|
v_cvt_f16_f32_e32 v96, v96 // 000000004A2C: 7EC01560
|
|
buffer_store_b16 v96, v153, s[16:19], 0 offen // 000000004A30: E0640000 80446099
|
|
v_mul_f32_e32 v97, v114, v97 // 000000004A38: 10C2C372
|
|
v_add_f32_e32 v76, v113, v97 // 000000004A3C: 0698C371
|
|
v_mov_b32_e32 v97, v76 // 000000004A40: 7EC2034C
|
|
v_cvt_f16_f32_e32 v97, v97 // 000000004A44: 7EC21561
|
|
buffer_store_b16 v97, v155, s[16:19], 0 offen // 000000004A48: E0640000 8044619B
|
|
v_mul_f32_e32 v98, v118, v98 // 000000004A50: 10C4C576
|
|
v_add_f32_e32 v76, v117, v98 // 000000004A54: 0698C575
|
|
v_mov_b32_e32 v98, v76 // 000000004A58: 7EC4034C
|
|
v_cvt_f16_f32_e32 v98, v98 // 000000004A5C: 7EC41562
|
|
buffer_store_b16 v98, v157, s[16:19], 0 offen // 000000004A60: E0640000 8044629D
|
|
v_mul_f32_e32 v99, v122, v99 // 000000004A68: 10C6C77A
|
|
v_add_f32_e32 v76, v121, v99 // 000000004A6C: 0698C779
|
|
v_mov_b32_e32 v99, v76 // 000000004A70: 7EC6034C
|
|
v_cvt_f16_f32_e32 v99, v99 // 000000004A74: 7EC61563
|
|
buffer_store_b16 v99, v159, s[16:19], 0 offen // 000000004A78: E0640000 8044639F
|
|
v_mul_f32_e32 v100, v114, v100 // 000000004A80: 10C8C972
|
|
v_add_f32_e32 v76, v113, v100 // 000000004A84: 0698C971
|
|
v_mov_b32_e32 v100, v76 // 000000004A88: 7EC8034C
|
|
v_cvt_f16_f32_e32 v100, v100 // 000000004A8C: 7EC81564
|
|
buffer_store_b16 v100, v161, s[16:19], 0 offen // 000000004A90: E0640000 804464A1
|
|
v_mul_f32_e32 v101, v118, v101 // 000000004A98: 10CACB76
|
|
v_add_f32_e32 v76, v117, v101 // 000000004A9C: 0698CB75
|
|
v_mov_b32_e32 v101, v76 // 000000004AA0: 7ECA034C
|
|
v_cvt_f16_f32_e32 v101, v101 // 000000004AA4: 7ECA1565
|
|
buffer_store_b16 v101, v163, s[16:19], 0 offen // 000000004AA8: E0640000 804465A3
|
|
v_mul_f32_e32 v102, v122, v102 // 000000004AB0: 10CCCD7A
|
|
v_add_f32_e32 v76, v121, v102 // 000000004AB4: 0698CD79
|
|
v_mov_b32_e32 v102, v76 // 000000004AB8: 7ECC034C
|
|
v_cvt_f16_f32_e32 v102, v102 // 000000004ABC: 7ECC1566
|
|
buffer_store_b16 v102, v165, s[16:19], 0 offen // 000000004AC0: E0640000 804466A5
|
|
v_mul_f32_e32 v103, v114, v103 // 000000004AC8: 10CECF72
|
|
v_add_f32_e32 v76, v113, v103 // 000000004ACC: 0698CF71
|
|
v_mov_b32_e32 v103, v76 // 000000004AD0: 7ECE034C
|
|
v_cvt_f16_f32_e32 v103, v103 // 000000004AD4: 7ECE1567
|
|
buffer_store_b16 v103, v167, s[16:19], 0 offen // 000000004AD8: E0640000 804467A7
|
|
v_mul_f32_e32 v104, v118, v104 // 000000004AE0: 10D0D176
|
|
v_add_f32_e32 v76, v117, v104 // 000000004AE4: 0698D175
|
|
v_mov_b32_e32 v104, v76 // 000000004AE8: 7ED0034C
|
|
v_cvt_f16_f32_e32 v104, v104 // 000000004AEC: 7ED01568
|
|
buffer_store_b16 v104, v169, s[16:19], 0 offen // 000000004AF0: E0640000 804468A9
|
|
v_mul_f32_e32 v105, v122, v105 // 000000004AF8: 10D2D37A
|
|
v_add_f32_e32 v76, v121, v105 // 000000004AFC: 0698D379
|
|
v_mov_b32_e32 v105, v76 // 000000004B00: 7ED2034C
|
|
v_cvt_f16_f32_e32 v105, v105 // 000000004B04: 7ED21569
|
|
buffer_store_b16 v105, v171, s[16:19], 0 offen // 000000004B08: E0640000 804469AB
|
|
v_mul_f32_e32 v106, v114, v106 // 000000004B10: 10D4D572
|
|
v_add_f32_e32 v76, v113, v106 // 000000004B14: 0698D571
|
|
v_mov_b32_e32 v106, v76 // 000000004B18: 7ED4034C
|
|
v_cvt_f16_f32_e32 v106, v106 // 000000004B1C: 7ED4156A
|
|
buffer_store_b16 v106, v173, s[16:19], 0 offen // 000000004B20: E0640000 80446AAD
|
|
v_mul_f32_e32 v107, v118, v107 // 000000004B28: 10D6D776
|
|
v_add_f32_e32 v76, v117, v107 // 000000004B2C: 0698D775
|
|
v_mov_b32_e32 v107, v76 // 000000004B30: 7ED6034C
|
|
v_cvt_f16_f32_e32 v107, v107 // 000000004B34: 7ED6156B
|
|
buffer_store_b16 v107, v175, s[16:19], 0 offen // 000000004B38: E0640000 80446BAF
|
|
v_mul_f32_e32 v108, v122, v108 // 000000004B40: 10D8D97A
|
|
v_add_f32_e32 v76, v121, v108 // 000000004B44: 0698D979
|
|
v_mov_b32_e32 v108, v76 // 000000004B48: 7ED8034C
|
|
v_cvt_f16_f32_e32 v108, v108 // 000000004B4C: 7ED8156C
|
|
buffer_store_b16 v108, v177, s[16:19], 0 offen // 000000004B50: E0640000 80446CB1
|
|
v_mul_f32_e32 v109, v114, v109 // 000000004B58: 10DADB72
|
|
v_add_f32_e32 v76, v113, v109 // 000000004B5C: 0698DB71
|
|
v_mov_b32_e32 v109, v76 // 000000004B60: 7EDA034C
|
|
v_cvt_f16_f32_e32 v109, v109 // 000000004B64: 7EDA156D
|
|
buffer_store_b16 v109, v179, s[16:19], 0 offen // 000000004B68: E0640000 80446DB3
|
|
v_mul_f32_e32 v110, v118, v110 // 000000004B70: 10DCDD76
|
|
v_add_f32_e32 v76, v117, v110 // 000000004B74: 0698DD75
|
|
v_mov_b32_e32 v110, v76 // 000000004B78: 7EDC034C
|
|
v_cvt_f16_f32_e32 v110, v110 // 000000004B7C: 7EDC156E
|
|
buffer_store_b16 v110, v181, s[16:19], 0 offen // 000000004B80: E0640000 80446EB5
|
|
v_mul_f32_e32 v111, v122, v111 // 000000004B88: 10DEDF7A
|
|
v_add_f32_e32 v76, v121, v111 // 000000004B8C: 0698DF79
|
|
v_mov_b32_e32 v111, v76 // 000000004B90: 7EDE034C
|
|
v_cvt_f16_f32_e32 v111, v111 // 000000004B94: 7EDE156F
|
|
buffer_store_b16 v111, v183, s[16:19], 0 offen // 000000004B98: E0640000 80446FB7
|
|
v_mul_f32_e32 v112, v114, v112 // 000000004BA0: 10E0E172
|
|
v_add_f32_e32 v76, v113, v112 // 000000004BA4: 0698E171
|
|
v_mov_b32_e32 v112, v76 // 000000004BA8: 7EE0034C
|
|
v_cvt_f16_f32_e32 v112, v112 // 000000004BAC: 7EE01570
|
|
buffer_store_b16 v112, v185, s[16:19], 0 offen // 000000004BB0: E0640000 804470B9
|
|
s_nop 0 // 000000004BB8: BF800000
|
|
v_mov_b32_e32 v78, 0x80000000 // 000000004BBC: 7E9C02FF 80000000
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000004BC4: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004BCC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004BD4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004BDC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004BE0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v116, v76, s32 // 000000004BE8: D5260074 0000414C
|
|
v_lshlrev_b32_e32 v116, 2, v116 // 000000004BF0: 30E8E882
|
|
ds_load_b32 v113, v116 // 000000004BF4: D8D80000 71000074
|
|
ds_load_b32 v114, v116 offset:512 // 000000004BFC: D8D80200 72000074
|
|
v_add_lshl_u32 v115, v75, v76, 1 // 000000004C04: D6470073 0206994B
|
|
v_cndmask_b32_e64 v115, v78, v115, s34 // 000000004C0C: D5010073 008AE74E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000004C14: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004C1C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004C24: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004C2C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004C30: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v120, v76, s32 // 000000004C38: D5260078 0000414C
|
|
v_lshlrev_b32_e32 v120, 2, v120 // 000000004C40: 30F0F082
|
|
ds_load_b32 v117, v120 // 000000004C44: D8D80000 75000078
|
|
ds_load_b32 v118, v120 offset:512 // 000000004C4C: D8D80200 76000078
|
|
v_add_lshl_u32 v119, v75, v76, 1 // 000000004C54: D6470077 0206994B
|
|
v_cndmask_b32_e64 v119, v78, v119, s34 // 000000004C5C: D5010077 008AEF4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000004C64: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 000000004C6C: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004C70: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000004C78: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 000000004C7C: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000004C84: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004C8C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004C94: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004C98: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v124, v72, s32 // 000000004CA0: D526007C 00004148
|
|
v_lshlrev_b32_e32 v124, 2, v124 // 000000004CA8: 30F8F882
|
|
ds_load_b32 v121, v124 // 000000004CAC: D8D80000 7900007C
|
|
ds_load_b32 v122, v124 offset:512 // 000000004CB4: D8D80200 7A00007C
|
|
v_add_lshl_u32 v123, v75, v72, 1 // 000000004CBC: D647007B 0206914B
|
|
v_cndmask_b32_e64 v123, v78, v123, s34 // 000000004CC4: D501007B 008AF74E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000004CCC: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004CD4: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004CDC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004CE4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004CE8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v126, v76, s32 // 000000004CF0: D526007E 0000414C
|
|
v_lshlrev_b32_e32 v126, 2, v126 // 000000004CF8: 30FCFC82
|
|
v_add_lshl_u32 v125, v75, v76, 1 // 000000004CFC: D647007D 0206994B
|
|
v_cndmask_b32_e64 v125, v78, v125, s34 // 000000004D04: D501007D 008AFB4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000004D0C: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004D14: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004D1C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004D24: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004D28: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v128, v76, s32 // 000000004D30: D5260080 0000414C
|
|
v_lshlrev_b32_e32 v128, 2, v128 // 000000004D38: 31010082
|
|
v_add_lshl_u32 v127, v75, v76, 1 // 000000004D3C: D647007F 0206994B
|
|
v_cndmask_b32_e64 v127, v78, v127, s34 // 000000004D44: D501007F 008AFF4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000004D4C: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 000000004D54: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004D58: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000004D60: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 000000004D64: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000004D6C: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004D74: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004D7C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004D80: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v130, v72, s32 // 000000004D88: D5260082 00004148
|
|
v_lshlrev_b32_e32 v130, 2, v130 // 000000004D90: 31050482
|
|
v_add_lshl_u32 v129, v75, v72, 1 // 000000004D94: D6470081 0206914B
|
|
v_cndmask_b32_e64 v129, v78, v129, s34 // 000000004D9C: D5010081 008B034E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000004DA4: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004DAC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004DB4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004DBC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004DC0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v132, v76, s32 // 000000004DC8: D5260084 0000414C
|
|
v_lshlrev_b32_e32 v132, 2, v132 // 000000004DD0: 31090882
|
|
v_add_lshl_u32 v131, v75, v76, 1 // 000000004DD4: D6470083 0206994B
|
|
v_cndmask_b32_e64 v131, v78, v131, s34 // 000000004DDC: D5010083 008B074E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000004DE4: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004DEC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004DF4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004DFC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004E00: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v134, v76, s32 // 000000004E08: D5260086 0000414C
|
|
v_lshlrev_b32_e32 v134, 2, v134 // 000000004E10: 310D0C82
|
|
v_add_lshl_u32 v133, v75, v76, 1 // 000000004E14: D6470085 0206994B
|
|
v_cndmask_b32_e64 v133, v78, v133, s34 // 000000004E1C: D5010085 008B0B4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000004E24: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 000000004E2C: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004E30: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000004E38: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 000000004E3C: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000004E44: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004E4C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004E54: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004E58: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v136, v72, s32 // 000000004E60: D5260088 00004148
|
|
v_lshlrev_b32_e32 v136, 2, v136 // 000000004E68: 31111082
|
|
v_add_lshl_u32 v135, v75, v72, 1 // 000000004E6C: D6470087 0206914B
|
|
v_cndmask_b32_e64 v135, v78, v135, s34 // 000000004E74: D5010087 008B0F4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000004E7C: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004E84: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004E8C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004E94: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004E98: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v138, v76, s32 // 000000004EA0: D526008A 0000414C
|
|
v_lshlrev_b32_e32 v138, 2, v138 // 000000004EA8: 31151482
|
|
v_add_lshl_u32 v137, v75, v76, 1 // 000000004EAC: D6470089 0206994B
|
|
v_cndmask_b32_e64 v137, v78, v137, s34 // 000000004EB4: D5010089 008B134E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000004EBC: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004EC4: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004ECC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004ED4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004ED8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v140, v76, s32 // 000000004EE0: D526008C 0000414C
|
|
v_lshlrev_b32_e32 v140, 2, v140 // 000000004EE8: 31191882
|
|
v_add_lshl_u32 v139, v75, v76, 1 // 000000004EEC: D647008B 0206994B
|
|
v_cndmask_b32_e64 v139, v78, v139, s34 // 000000004EF4: D501008B 008B174E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000004EFC: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 000000004F04: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004F08: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000004F10: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 000000004F14: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000004F1C: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004F24: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004F2C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004F30: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v142, v72, s32 // 000000004F38: D526008E 00004148
|
|
v_lshlrev_b32_e32 v142, 2, v142 // 000000004F40: 311D1C82
|
|
v_add_lshl_u32 v141, v75, v72, 1 // 000000004F44: D647008D 0206914B
|
|
v_cndmask_b32_e64 v141, v78, v141, s34 // 000000004F4C: D501008D 008B1B4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000004F54: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004F5C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004F64: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004F6C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004F70: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v144, v76, s32 // 000000004F78: D5260090 0000414C
|
|
v_lshlrev_b32_e32 v144, 2, v144 // 000000004F80: 31212082
|
|
v_add_lshl_u32 v143, v75, v76, 1 // 000000004F84: D647008F 0206994B
|
|
v_cndmask_b32_e64 v143, v78, v143, s34 // 000000004F8C: D501008F 008B1F4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000004F94: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000004F9C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004FA4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000004FAC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000004FB0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v146, v76, s32 // 000000004FB8: D5260092 0000414C
|
|
v_lshlrev_b32_e32 v146, 2, v146 // 000000004FC0: 31252482
|
|
v_add_lshl_u32 v145, v75, v76, 1 // 000000004FC4: D6470091 0206994B
|
|
v_cndmask_b32_e64 v145, v78, v145, s34 // 000000004FCC: D5010091 008B234E
|
|
v_add_co_u32 v73, vcc_lo, v73, 18 // 000000004FD4: D7006A49 00012549
|
|
s_mul_i32 s32, s38, 18 // 000000004FDC: 96209226
|
|
v_add_nc_i32 v74, v74, s32 // 000000004FE0: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 18 // 000000004FE8: 96209224
|
|
v_add_nc_i32 v75, v75, s32 // 000000004FEC: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000004FF4: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000004FFC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000005004: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005008: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v148, v72, s32 // 000000005010: D5260094 00004148
|
|
v_lshlrev_b32_e32 v148, 2, v148 // 000000005018: 31292882
|
|
v_add_lshl_u32 v147, v75, v72, 1 // 00000000501C: D6470093 0206914B
|
|
v_cndmask_b32_e64 v147, v78, v147, s34 // 000000005024: D5010093 008B274E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 00000000502C: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000005034: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000503C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000005044: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005048: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v150, v76, s32 // 000000005050: D5260096 0000414C
|
|
v_lshlrev_b32_e32 v150, 2, v150 // 000000005058: 312D2C82
|
|
v_add_lshl_u32 v149, v75, v76, 1 // 00000000505C: D6470095 0206994B
|
|
v_cndmask_b32_e64 v149, v78, v149, s34 // 000000005064: D5010095 008B2B4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 00000000506C: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000005074: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000507C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000005084: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005088: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v152, v76, s32 // 000000005090: D5260098 0000414C
|
|
v_lshlrev_b32_e32 v152, 2, v152 // 000000005098: 31313082
|
|
v_add_lshl_u32 v151, v75, v76, 1 // 00000000509C: D6470097 0206994B
|
|
v_cndmask_b32_e64 v151, v78, v151, s34 // 0000000050A4: D5010097 008B2F4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 0000000050AC: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 0000000050B4: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 0000000050B8: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 0000000050C0: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 0000000050C4: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 0000000050CC: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000050D4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000050DC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000050E0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v154, v72, s32 // 0000000050E8: D526009A 00004148
|
|
v_lshlrev_b32_e32 v154, 2, v154 // 0000000050F0: 31353482
|
|
v_add_lshl_u32 v153, v75, v72, 1 // 0000000050F4: D6470099 0206914B
|
|
v_cndmask_b32_e64 v153, v78, v153, s34 // 0000000050FC: D5010099 008B334E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000005104: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000510C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000005114: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000511C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005120: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v156, v76, s32 // 000000005128: D526009C 0000414C
|
|
v_lshlrev_b32_e32 v156, 2, v156 // 000000005130: 31393882
|
|
v_add_lshl_u32 v155, v75, v76, 1 // 000000005134: D647009B 0206994B
|
|
v_cndmask_b32_e64 v155, v78, v155, s34 // 00000000513C: D501009B 008B374E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000005144: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000514C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000005154: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000515C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005160: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v158, v76, s32 // 000000005168: D526009E 0000414C
|
|
v_lshlrev_b32_e32 v158, 2, v158 // 000000005170: 313D3C82
|
|
v_add_lshl_u32 v157, v75, v76, 1 // 000000005174: D647009D 0206994B
|
|
v_cndmask_b32_e64 v157, v78, v157, s34 // 00000000517C: D501009D 008B3B4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000005184: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 00000000518C: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000005190: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000005198: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 00000000519C: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 0000000051A4: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000051AC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000051B4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000051B8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v160, v72, s32 // 0000000051C0: D52600A0 00004148
|
|
v_lshlrev_b32_e32 v160, 2, v160 // 0000000051C8: 31414082
|
|
v_add_lshl_u32 v159, v75, v72, 1 // 0000000051CC: D647009F 0206914B
|
|
v_cndmask_b32_e64 v159, v78, v159, s34 // 0000000051D4: D501009F 008B3F4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 0000000051DC: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000051E4: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000051EC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000051F4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000051F8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v162, v76, s32 // 000000005200: D52600A2 0000414C
|
|
v_lshlrev_b32_e32 v162, 2, v162 // 000000005208: 31454482
|
|
v_add_lshl_u32 v161, v75, v76, 1 // 00000000520C: D64700A1 0206994B
|
|
v_cndmask_b32_e64 v161, v78, v161, s34 // 000000005214: D50100A1 008B434E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 00000000521C: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000005224: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000522C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000005234: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005238: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v164, v76, s32 // 000000005240: D52600A4 0000414C
|
|
v_lshlrev_b32_e32 v164, 2, v164 // 000000005248: 31494882
|
|
v_add_lshl_u32 v163, v75, v76, 1 // 00000000524C: D64700A3 0206994B
|
|
v_cndmask_b32_e64 v163, v78, v163, s34 // 000000005254: D50100A3 008B474E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 00000000525C: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 000000005264: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000005268: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000005270: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 000000005274: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 00000000527C: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000005284: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000528C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005290: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v166, v72, s32 // 000000005298: D52600A6 00004148
|
|
v_lshlrev_b32_e32 v166, 2, v166 // 0000000052A0: 314D4C82
|
|
v_add_lshl_u32 v165, v75, v72, 1 // 0000000052A4: D64700A5 0206914B
|
|
v_cndmask_b32_e64 v165, v78, v165, s34 // 0000000052AC: D50100A5 008B4B4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 0000000052B4: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000052BC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000052C4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000052CC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000052D0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v168, v76, s32 // 0000000052D8: D52600A8 0000414C
|
|
v_lshlrev_b32_e32 v168, 2, v168 // 0000000052E0: 31515082
|
|
v_add_lshl_u32 v167, v75, v76, 1 // 0000000052E4: D64700A7 0206994B
|
|
v_cndmask_b32_e64 v167, v78, v167, s34 // 0000000052EC: D50100A7 008B4F4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 0000000052F4: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000052FC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000005304: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000530C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005310: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v170, v76, s32 // 000000005318: D52600AA 0000414C
|
|
v_lshlrev_b32_e32 v170, 2, v170 // 000000005320: 31555482
|
|
v_add_lshl_u32 v169, v75, v76, 1 // 000000005324: D64700A9 0206994B
|
|
v_cndmask_b32_e64 v169, v78, v169, s34 // 00000000532C: D50100A9 008B534E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000005334: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 00000000533C: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000005340: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000005348: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 00000000534C: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000005354: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000535C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000005364: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005368: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v172, v72, s32 // 000000005370: D52600AC 00004148
|
|
v_lshlrev_b32_e32 v172, 2, v172 // 000000005378: 31595882
|
|
v_add_lshl_u32 v171, v75, v72, 1 // 00000000537C: D64700AB 0206914B
|
|
v_cndmask_b32_e64 v171, v78, v171, s34 // 000000005384: D50100AB 008B574E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 00000000538C: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000005394: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000539C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000053A4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000053A8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v174, v76, s32 // 0000000053B0: D52600AE 0000414C
|
|
v_lshlrev_b32_e32 v174, 2, v174 // 0000000053B8: 315D5C82
|
|
v_add_lshl_u32 v173, v75, v76, 1 // 0000000053BC: D64700AD 0206994B
|
|
v_cndmask_b32_e64 v173, v78, v173, s34 // 0000000053C4: D50100AD 008B5B4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 0000000053CC: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000053D4: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000053DC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000053E4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000053E8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v176, v76, s32 // 0000000053F0: D52600B0 0000414C
|
|
v_lshlrev_b32_e32 v176, 2, v176 // 0000000053F8: 31616082
|
|
v_add_lshl_u32 v175, v75, v76, 1 // 0000000053FC: D64700AF 0206994B
|
|
v_cndmask_b32_e64 v175, v78, v175, s34 // 000000005404: D50100AF 008B5F4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 00000000540C: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 000000005414: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 000000005418: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 000000005420: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 000000005424: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 00000000542C: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000005434: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000543C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005440: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v178, v72, s32 // 000000005448: D52600B2 00004148
|
|
v_lshlrev_b32_e32 v178, 2, v178 // 000000005450: 31656482
|
|
v_add_lshl_u32 v177, v75, v72, 1 // 000000005454: D64700B1 0206914B
|
|
v_cndmask_b32_e64 v177, v78, v177, s34 // 00000000545C: D50100B1 008B634E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 000000005464: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000546C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000005474: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000547C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005480: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v180, v76, s32 // 000000005488: D52600B4 0000414C
|
|
v_lshlrev_b32_e32 v180, 2, v180 // 000000005490: 31696882
|
|
v_add_lshl_u32 v179, v75, v76, 1 // 000000005494: D64700B3 0206994B
|
|
v_cndmask_b32_e64 v179, v78, v179, s34 // 00000000549C: D50100B3 008B674E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 0000000054A4: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 0000000054AC: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000054B4: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000054BC: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000054C0: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v182, v76, s32 // 0000000054C8: D52600B6 0000414C
|
|
v_lshlrev_b32_e32 v182, 2, v182 // 0000000054D0: 316D6C82
|
|
v_add_lshl_u32 v181, v75, v76, 1 // 0000000054D4: D64700B5 0206994B
|
|
v_cndmask_b32_e64 v181, v78, v181, s34 // 0000000054DC: D50100B5 008B6B4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 0000000054E4: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 0000000054EC: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 0000000054F0: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 0000000054F8: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 0000000054FC: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 000000005504: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000550C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000005514: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005518: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v184, v72, s32 // 000000005520: D52600B8 00004148
|
|
v_lshlrev_b32_e32 v184, 2, v184 // 000000005528: 31717082
|
|
v_add_lshl_u32 v183, v75, v72, 1 // 00000000552C: D64700B7 0206914B
|
|
v_cndmask_b32_e64 v183, v78, v183, s34 // 000000005534: D50100B7 008B6F4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 00000000553C: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000005544: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 00000000554C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000005554: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005558: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v186, v76, s32 // 000000005560: D52600BA 0000414C
|
|
v_lshlrev_b32_e32 v186, 2, v186 // 000000005568: 31757482
|
|
v_add_lshl_u32 v185, v75, v76, 1 // 00000000556C: D64700B9 0206994B
|
|
v_cndmask_b32_e64 v185, v78, v185, s34 // 000000005574: D50100B9 008B734E
|
|
v_mov_b32_e32 v79, v35 // 00000000557C: 7E9E0323
|
|
v_mov_b32_e32 v80, v43 // 000000005580: 7EA0032B
|
|
v_mov_b32_e32 v81, v28 // 000000005584: 7EA2031C
|
|
v_mov_b32_e32 v82, v36 // 000000005588: 7EA40324
|
|
v_mov_b32_e32 v83, v44 // 00000000558C: 7EA6032C
|
|
v_mov_b32_e32 v84, v29 // 000000005590: 7EA8031D
|
|
v_mov_b32_e32 v85, v37 // 000000005594: 7EAA0325
|
|
v_mov_b32_e32 v86, v45 // 000000005598: 7EAC032D
|
|
v_mov_b32_e32 v87, v30 // 00000000559C: 7EAE031E
|
|
v_mov_b32_e32 v88, v38 // 0000000055A0: 7EB00326
|
|
v_mov_b32_e32 v89, v46 // 0000000055A4: 7EB2032E
|
|
v_mov_b32_e32 v90, v31 // 0000000055A8: 7EB4031F
|
|
v_mov_b32_e32 v91, v39 // 0000000055AC: 7EB60327
|
|
v_mov_b32_e32 v92, v47 // 0000000055B0: 7EB8032F
|
|
v_mov_b32_e32 v93, v48 // 0000000055B4: 7EBA0330
|
|
v_mov_b32_e32 v94, v56 // 0000000055B8: 7EBC0338
|
|
v_mov_b32_e32 v95, v64 // 0000000055BC: 7EBE0340
|
|
v_mov_b32_e32 v96, v49 // 0000000055C0: 7EC00331
|
|
v_mov_b32_e32 v97, v57 // 0000000055C4: 7EC20339
|
|
v_mov_b32_e32 v98, v65 // 0000000055C8: 7EC40341
|
|
v_mov_b32_e32 v99, v50 // 0000000055CC: 7EC60332
|
|
v_mov_b32_e32 v100, v58 // 0000000055D0: 7EC8033A
|
|
v_mov_b32_e32 v101, v66 // 0000000055D4: 7ECA0342
|
|
v_mov_b32_e32 v102, v51 // 0000000055D8: 7ECC0333
|
|
v_mov_b32_e32 v103, v59 // 0000000055DC: 7ECE033B
|
|
v_mov_b32_e32 v104, v67 // 0000000055E0: 7ED00343
|
|
v_mov_b32_e32 v105, v52 // 0000000055E4: 7ED20334
|
|
v_mov_b32_e32 v106, v60 // 0000000055E8: 7ED4033C
|
|
v_mov_b32_e32 v107, v68 // 0000000055EC: 7ED60344
|
|
v_mov_b32_e32 v108, v53 // 0000000055F0: 7ED80335
|
|
v_mov_b32_e32 v109, v61 // 0000000055F4: 7EDA033D
|
|
v_mov_b32_e32 v110, v69 // 0000000055F8: 7EDC0345
|
|
v_mov_b32_e32 v111, v54 // 0000000055FC: 7EDE0336
|
|
v_mov_b32_e32 v112, v62 // 000000005600: 7EE0033E
|
|
s_waitcnt lgkmcnt(0) // 000000005604: BF89FC07
|
|
v_mul_f32_e32 v79, v114, v79 // 000000005608: 109E9F72
|
|
v_add_f32_e32 v76, v113, v79 // 00000000560C: 06989F71
|
|
v_mov_b32_e32 v79, v76 // 000000005610: 7E9E034C
|
|
v_cvt_f16_f32_e32 v79, v79 // 000000005614: 7E9E154F
|
|
buffer_store_b16 v79, v115, s[16:19], 0 offen // 000000005618: E0640000 80444F73
|
|
v_mul_f32_e32 v80, v118, v80 // 000000005620: 10A0A176
|
|
v_add_f32_e32 v76, v117, v80 // 000000005624: 0698A175
|
|
v_mov_b32_e32 v80, v76 // 000000005628: 7EA0034C
|
|
v_cvt_f16_f32_e32 v80, v80 // 00000000562C: 7EA01550
|
|
buffer_store_b16 v80, v119, s[16:19], 0 offen // 000000005630: E0640000 80445077
|
|
v_mul_f32_e32 v81, v122, v81 // 000000005638: 10A2A37A
|
|
v_add_f32_e32 v76, v121, v81 // 00000000563C: 0698A379
|
|
v_mov_b32_e32 v81, v76 // 000000005640: 7EA2034C
|
|
v_cvt_f16_f32_e32 v81, v81 // 000000005644: 7EA21551
|
|
buffer_store_b16 v81, v123, s[16:19], 0 offen // 000000005648: E0640000 8044517B
|
|
v_mul_f32_e32 v82, v114, v82 // 000000005650: 10A4A572
|
|
v_add_f32_e32 v76, v113, v82 // 000000005654: 0698A571
|
|
v_mov_b32_e32 v82, v76 // 000000005658: 7EA4034C
|
|
v_cvt_f16_f32_e32 v82, v82 // 00000000565C: 7EA41552
|
|
buffer_store_b16 v82, v125, s[16:19], 0 offen // 000000005660: E0640000 8044527D
|
|
v_mul_f32_e32 v83, v118, v83 // 000000005668: 10A6A776
|
|
v_add_f32_e32 v76, v117, v83 // 00000000566C: 0698A775
|
|
v_mov_b32_e32 v83, v76 // 000000005670: 7EA6034C
|
|
v_cvt_f16_f32_e32 v83, v83 // 000000005674: 7EA61553
|
|
buffer_store_b16 v83, v127, s[16:19], 0 offen // 000000005678: E0640000 8044537F
|
|
v_mul_f32_e32 v84, v122, v84 // 000000005680: 10A8A97A
|
|
v_add_f32_e32 v76, v121, v84 // 000000005684: 0698A979
|
|
v_mov_b32_e32 v84, v76 // 000000005688: 7EA8034C
|
|
v_cvt_f16_f32_e32 v84, v84 // 00000000568C: 7EA81554
|
|
buffer_store_b16 v84, v129, s[16:19], 0 offen // 000000005690: E0640000 80445481
|
|
v_mul_f32_e32 v85, v114, v85 // 000000005698: 10AAAB72
|
|
v_add_f32_e32 v76, v113, v85 // 00000000569C: 0698AB71
|
|
v_mov_b32_e32 v85, v76 // 0000000056A0: 7EAA034C
|
|
v_cvt_f16_f32_e32 v85, v85 // 0000000056A4: 7EAA1555
|
|
buffer_store_b16 v85, v131, s[16:19], 0 offen // 0000000056A8: E0640000 80445583
|
|
v_mul_f32_e32 v86, v118, v86 // 0000000056B0: 10ACAD76
|
|
v_add_f32_e32 v76, v117, v86 // 0000000056B4: 0698AD75
|
|
v_mov_b32_e32 v86, v76 // 0000000056B8: 7EAC034C
|
|
v_cvt_f16_f32_e32 v86, v86 // 0000000056BC: 7EAC1556
|
|
buffer_store_b16 v86, v133, s[16:19], 0 offen // 0000000056C0: E0640000 80445685
|
|
v_mul_f32_e32 v87, v122, v87 // 0000000056C8: 10AEAF7A
|
|
v_add_f32_e32 v76, v121, v87 // 0000000056CC: 0698AF79
|
|
v_mov_b32_e32 v87, v76 // 0000000056D0: 7EAE034C
|
|
v_cvt_f16_f32_e32 v87, v87 // 0000000056D4: 7EAE1557
|
|
buffer_store_b16 v87, v135, s[16:19], 0 offen // 0000000056D8: E0640000 80445787
|
|
v_mul_f32_e32 v88, v114, v88 // 0000000056E0: 10B0B172
|
|
v_add_f32_e32 v76, v113, v88 // 0000000056E4: 0698B171
|
|
v_mov_b32_e32 v88, v76 // 0000000056E8: 7EB0034C
|
|
v_cvt_f16_f32_e32 v88, v88 // 0000000056EC: 7EB01558
|
|
buffer_store_b16 v88, v137, s[16:19], 0 offen // 0000000056F0: E0640000 80445889
|
|
v_mul_f32_e32 v89, v118, v89 // 0000000056F8: 10B2B376
|
|
v_add_f32_e32 v76, v117, v89 // 0000000056FC: 0698B375
|
|
v_mov_b32_e32 v89, v76 // 000000005700: 7EB2034C
|
|
v_cvt_f16_f32_e32 v89, v89 // 000000005704: 7EB21559
|
|
buffer_store_b16 v89, v139, s[16:19], 0 offen // 000000005708: E0640000 8044598B
|
|
v_mul_f32_e32 v90, v122, v90 // 000000005710: 10B4B57A
|
|
v_add_f32_e32 v76, v121, v90 // 000000005714: 0698B579
|
|
v_mov_b32_e32 v90, v76 // 000000005718: 7EB4034C
|
|
v_cvt_f16_f32_e32 v90, v90 // 00000000571C: 7EB4155A
|
|
buffer_store_b16 v90, v141, s[16:19], 0 offen // 000000005720: E0640000 80445A8D
|
|
v_mul_f32_e32 v91, v114, v91 // 000000005728: 10B6B772
|
|
v_add_f32_e32 v76, v113, v91 // 00000000572C: 0698B771
|
|
v_mov_b32_e32 v91, v76 // 000000005730: 7EB6034C
|
|
v_cvt_f16_f32_e32 v91, v91 // 000000005734: 7EB6155B
|
|
buffer_store_b16 v91, v143, s[16:19], 0 offen // 000000005738: E0640000 80445B8F
|
|
v_mul_f32_e32 v92, v118, v92 // 000000005740: 10B8B976
|
|
v_add_f32_e32 v76, v117, v92 // 000000005744: 0698B975
|
|
v_mov_b32_e32 v92, v76 // 000000005748: 7EB8034C
|
|
v_cvt_f16_f32_e32 v92, v92 // 00000000574C: 7EB8155C
|
|
buffer_store_b16 v92, v145, s[16:19], 0 offen // 000000005750: E0640000 80445C91
|
|
v_mul_f32_e32 v93, v122, v93 // 000000005758: 10BABB7A
|
|
v_add_f32_e32 v76, v121, v93 // 00000000575C: 0698BB79
|
|
v_mov_b32_e32 v93, v76 // 000000005760: 7EBA034C
|
|
v_cvt_f16_f32_e32 v93, v93 // 000000005764: 7EBA155D
|
|
buffer_store_b16 v93, v147, s[16:19], 0 offen // 000000005768: E0640000 80445D93
|
|
v_mul_f32_e32 v94, v114, v94 // 000000005770: 10BCBD72
|
|
v_add_f32_e32 v76, v113, v94 // 000000005774: 0698BD71
|
|
v_mov_b32_e32 v94, v76 // 000000005778: 7EBC034C
|
|
v_cvt_f16_f32_e32 v94, v94 // 00000000577C: 7EBC155E
|
|
buffer_store_b16 v94, v149, s[16:19], 0 offen // 000000005780: E0640000 80445E95
|
|
v_mul_f32_e32 v95, v118, v95 // 000000005788: 10BEBF76
|
|
v_add_f32_e32 v76, v117, v95 // 00000000578C: 0698BF75
|
|
v_mov_b32_e32 v95, v76 // 000000005790: 7EBE034C
|
|
v_cvt_f16_f32_e32 v95, v95 // 000000005794: 7EBE155F
|
|
buffer_store_b16 v95, v151, s[16:19], 0 offen // 000000005798: E0640000 80445F97
|
|
v_mul_f32_e32 v96, v122, v96 // 0000000057A0: 10C0C17A
|
|
v_add_f32_e32 v76, v121, v96 // 0000000057A4: 0698C179
|
|
v_mov_b32_e32 v96, v76 // 0000000057A8: 7EC0034C
|
|
v_cvt_f16_f32_e32 v96, v96 // 0000000057AC: 7EC01560
|
|
buffer_store_b16 v96, v153, s[16:19], 0 offen // 0000000057B0: E0640000 80446099
|
|
v_mul_f32_e32 v97, v114, v97 // 0000000057B8: 10C2C372
|
|
v_add_f32_e32 v76, v113, v97 // 0000000057BC: 0698C371
|
|
v_mov_b32_e32 v97, v76 // 0000000057C0: 7EC2034C
|
|
v_cvt_f16_f32_e32 v97, v97 // 0000000057C4: 7EC21561
|
|
buffer_store_b16 v97, v155, s[16:19], 0 offen // 0000000057C8: E0640000 8044619B
|
|
v_mul_f32_e32 v98, v118, v98 // 0000000057D0: 10C4C576
|
|
v_add_f32_e32 v76, v117, v98 // 0000000057D4: 0698C575
|
|
v_mov_b32_e32 v98, v76 // 0000000057D8: 7EC4034C
|
|
v_cvt_f16_f32_e32 v98, v98 // 0000000057DC: 7EC41562
|
|
buffer_store_b16 v98, v157, s[16:19], 0 offen // 0000000057E0: E0640000 8044629D
|
|
v_mul_f32_e32 v99, v122, v99 // 0000000057E8: 10C6C77A
|
|
v_add_f32_e32 v76, v121, v99 // 0000000057EC: 0698C779
|
|
v_mov_b32_e32 v99, v76 // 0000000057F0: 7EC6034C
|
|
v_cvt_f16_f32_e32 v99, v99 // 0000000057F4: 7EC61563
|
|
buffer_store_b16 v99, v159, s[16:19], 0 offen // 0000000057F8: E0640000 8044639F
|
|
v_mul_f32_e32 v100, v114, v100 // 000000005800: 10C8C972
|
|
v_add_f32_e32 v76, v113, v100 // 000000005804: 0698C971
|
|
v_mov_b32_e32 v100, v76 // 000000005808: 7EC8034C
|
|
v_cvt_f16_f32_e32 v100, v100 // 00000000580C: 7EC81564
|
|
buffer_store_b16 v100, v161, s[16:19], 0 offen // 000000005810: E0640000 804464A1
|
|
v_mul_f32_e32 v101, v118, v101 // 000000005818: 10CACB76
|
|
v_add_f32_e32 v76, v117, v101 // 00000000581C: 0698CB75
|
|
v_mov_b32_e32 v101, v76 // 000000005820: 7ECA034C
|
|
v_cvt_f16_f32_e32 v101, v101 // 000000005824: 7ECA1565
|
|
buffer_store_b16 v101, v163, s[16:19], 0 offen // 000000005828: E0640000 804465A3
|
|
v_mul_f32_e32 v102, v122, v102 // 000000005830: 10CCCD7A
|
|
v_add_f32_e32 v76, v121, v102 // 000000005834: 0698CD79
|
|
v_mov_b32_e32 v102, v76 // 000000005838: 7ECC034C
|
|
v_cvt_f16_f32_e32 v102, v102 // 00000000583C: 7ECC1566
|
|
buffer_store_b16 v102, v165, s[16:19], 0 offen // 000000005840: E0640000 804466A5
|
|
v_mul_f32_e32 v103, v114, v103 // 000000005848: 10CECF72
|
|
v_add_f32_e32 v76, v113, v103 // 00000000584C: 0698CF71
|
|
v_mov_b32_e32 v103, v76 // 000000005850: 7ECE034C
|
|
v_cvt_f16_f32_e32 v103, v103 // 000000005854: 7ECE1567
|
|
buffer_store_b16 v103, v167, s[16:19], 0 offen // 000000005858: E0640000 804467A7
|
|
v_mul_f32_e32 v104, v118, v104 // 000000005860: 10D0D176
|
|
v_add_f32_e32 v76, v117, v104 // 000000005864: 0698D175
|
|
v_mov_b32_e32 v104, v76 // 000000005868: 7ED0034C
|
|
v_cvt_f16_f32_e32 v104, v104 // 00000000586C: 7ED01568
|
|
buffer_store_b16 v104, v169, s[16:19], 0 offen // 000000005870: E0640000 804468A9
|
|
v_mul_f32_e32 v105, v122, v105 // 000000005878: 10D2D37A
|
|
v_add_f32_e32 v76, v121, v105 // 00000000587C: 0698D379
|
|
v_mov_b32_e32 v105, v76 // 000000005880: 7ED2034C
|
|
v_cvt_f16_f32_e32 v105, v105 // 000000005884: 7ED21569
|
|
buffer_store_b16 v105, v171, s[16:19], 0 offen // 000000005888: E0640000 804469AB
|
|
v_mul_f32_e32 v106, v114, v106 // 000000005890: 10D4D572
|
|
v_add_f32_e32 v76, v113, v106 // 000000005894: 0698D571
|
|
v_mov_b32_e32 v106, v76 // 000000005898: 7ED4034C
|
|
v_cvt_f16_f32_e32 v106, v106 // 00000000589C: 7ED4156A
|
|
buffer_store_b16 v106, v173, s[16:19], 0 offen // 0000000058A0: E0640000 80446AAD
|
|
v_mul_f32_e32 v107, v118, v107 // 0000000058A8: 10D6D776
|
|
v_add_f32_e32 v76, v117, v107 // 0000000058AC: 0698D775
|
|
v_mov_b32_e32 v107, v76 // 0000000058B0: 7ED6034C
|
|
v_cvt_f16_f32_e32 v107, v107 // 0000000058B4: 7ED6156B
|
|
buffer_store_b16 v107, v175, s[16:19], 0 offen // 0000000058B8: E0640000 80446BAF
|
|
v_mul_f32_e32 v108, v122, v108 // 0000000058C0: 10D8D97A
|
|
v_add_f32_e32 v76, v121, v108 // 0000000058C4: 0698D979
|
|
v_mov_b32_e32 v108, v76 // 0000000058C8: 7ED8034C
|
|
v_cvt_f16_f32_e32 v108, v108 // 0000000058CC: 7ED8156C
|
|
buffer_store_b16 v108, v177, s[16:19], 0 offen // 0000000058D0: E0640000 80446CB1
|
|
v_mul_f32_e32 v109, v114, v109 // 0000000058D8: 10DADB72
|
|
v_add_f32_e32 v76, v113, v109 // 0000000058DC: 0698DB71
|
|
v_mov_b32_e32 v109, v76 // 0000000058E0: 7EDA034C
|
|
v_cvt_f16_f32_e32 v109, v109 // 0000000058E4: 7EDA156D
|
|
buffer_store_b16 v109, v179, s[16:19], 0 offen // 0000000058E8: E0640000 80446DB3
|
|
v_mul_f32_e32 v110, v118, v110 // 0000000058F0: 10DCDD76
|
|
v_add_f32_e32 v76, v117, v110 // 0000000058F4: 0698DD75
|
|
v_mov_b32_e32 v110, v76 // 0000000058F8: 7EDC034C
|
|
v_cvt_f16_f32_e32 v110, v110 // 0000000058FC: 7EDC156E
|
|
buffer_store_b16 v110, v181, s[16:19], 0 offen // 000000005900: E0640000 80446EB5
|
|
v_mul_f32_e32 v111, v122, v111 // 000000005908: 10DEDF7A
|
|
v_add_f32_e32 v76, v121, v111 // 00000000590C: 0698DF79
|
|
v_mov_b32_e32 v111, v76 // 000000005910: 7EDE034C
|
|
v_cvt_f16_f32_e32 v111, v111 // 000000005914: 7EDE156F
|
|
buffer_store_b16 v111, v183, s[16:19], 0 offen // 000000005918: E0640000 80446FB7
|
|
v_mul_f32_e32 v112, v114, v112 // 000000005920: 10E0E172
|
|
v_add_f32_e32 v76, v113, v112 // 000000005924: 0698E171
|
|
v_mov_b32_e32 v112, v76 // 000000005928: 7EE0034C
|
|
v_cvt_f16_f32_e32 v112, v112 // 00000000592C: 7EE01570
|
|
buffer_store_b16 v112, v185, s[16:19], 0 offen // 000000005930: E0640000 804470B9
|
|
s_nop 0 // 000000005938: BF800000
|
|
v_mov_b32_e32 v78, 0x80000000 // 00000000593C: 7E9C02FF 80000000
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000005944: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 00000000594C: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000005954: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 00000000595C: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005960: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v86, v76, s32 // 000000005968: D5260056 0000414C
|
|
v_lshlrev_b32_e32 v86, 2, v86 // 000000005970: 30ACAC82
|
|
ds_load_b32 v83, v86 // 000000005974: D8D80000 53000056
|
|
ds_load_b32 v84, v86 offset:512 // 00000000597C: D8D80200 54000056
|
|
v_add_lshl_u32 v85, v75, v76, 1 // 000000005984: D6470055 0206994B
|
|
v_cndmask_b32_e64 v85, v78, v85, s34 // 00000000598C: D5010055 008AAB4E
|
|
v_add_co_u32 v73, vcc_lo, v73, 2 // 000000005994: D7006A49 00010549
|
|
s_mul_i32 s32, s38, 2 // 00000000599C: 96208226
|
|
v_add_nc_i32 v74, v74, s32 // 0000000059A0: D726004A 0000414A
|
|
s_mul_i32 s32, s36, 2 // 0000000059A8: 96208224
|
|
v_add_nc_i32 v75, v75, s32 // 0000000059AC: D726004B 0000414B
|
|
v_cmp_lt_u32_e64 s32, v72, s24 // 0000000059B4: D4490020 00003148
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 0000000059BC: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 0000000059C4: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 0000000059C8: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v90, v72, s32 // 0000000059D0: D526005A 00004148
|
|
v_lshlrev_b32_e32 v90, 2, v90 // 0000000059D8: 30B4B482
|
|
ds_load_b32 v87, v90 // 0000000059DC: D8D80000 5700005A
|
|
ds_load_b32 v88, v90 offset:512 // 0000000059E4: D8D80200 5800005A
|
|
v_add_lshl_u32 v89, v75, v72, 1 // 0000000059EC: D6470059 0206914B
|
|
v_cndmask_b32_e64 v89, v78, v89, s34 // 0000000059F4: D5010059 008AB34E
|
|
v_add_co_u32 v76, vcc_lo, v72, 32 // 0000000059FC: D7006A4C 00014148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000005A04: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000005A0C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000005A14: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005A18: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v94, v76, s32 // 000000005A20: D526005E 0000414C
|
|
v_lshlrev_b32_e32 v94, 2, v94 // 000000005A28: 30BCBC82
|
|
ds_load_b32 v91, v94 // 000000005A2C: D8D80000 5B00005E
|
|
ds_load_b32 v92, v94 offset:512 // 000000005A34: D8D80200 5C00005E
|
|
v_add_lshl_u32 v93, v75, v76, 1 // 000000005A3C: D647005D 0206994B
|
|
v_cndmask_b32_e64 v93, v78, v93, s34 // 000000005A44: D501005D 008ABB4E
|
|
v_add_co_u32 v76, vcc_lo, v72, 64 // 000000005A4C: D7006A4C 00018148
|
|
v_cmp_lt_u32_e64 s32, v76, s24 // 000000005A54: D4490020 0000314C
|
|
v_cmp_lt_u32_e64 s34, v73, s25 // 000000005A5C: D4490022 00003349
|
|
s_and_b32 s34, s32, s34 // 000000005A64: 8B222220
|
|
s_mul_i32 s32, 0x60, s2 // 000000005A68: 962002FF 00000060
|
|
v_sub_nc_u32_e64 v96, v76, s32 // 000000005A70: D5260060 0000414C
|
|
v_lshlrev_b32_e32 v96, 2, v96 // 000000005A78: 30C0C082
|
|
v_add_lshl_u32 v95, v75, v76, 1 // 000000005A7C: D647005F 0206994B
|
|
v_cndmask_b32_e64 v95, v78, v95, s34 // 000000005A84: D501005F 008ABF4E
|
|
v_mov_b32_e32 v79, v70 // 000000005A8C: 7E9E0346
|
|
v_mov_b32_e32 v80, v55 // 000000005A90: 7EA00337
|
|
v_mov_b32_e32 v81, v63 // 000000005A94: 7EA2033F
|
|
v_mov_b32_e32 v82, v71 // 000000005A98: 7EA40347
|
|
s_waitcnt lgkmcnt(0) // 000000005A9C: BF89FC07
|
|
v_mul_f32_e32 v79, v84, v79 // 000000005AA0: 109E9F54
|
|
v_add_f32_e32 v76, v83, v79 // 000000005AA4: 06989F53
|
|
v_mov_b32_e32 v79, v76 // 000000005AA8: 7E9E034C
|
|
v_cvt_f16_f32_e32 v79, v79 // 000000005AAC: 7E9E154F
|
|
buffer_store_b16 v79, v85, s[16:19], 0 offen // 000000005AB0: E0640000 80444F55
|
|
v_mul_f32_e32 v80, v88, v80 // 000000005AB8: 10A0A158
|
|
v_add_f32_e32 v76, v87, v80 // 000000005ABC: 0698A157
|
|
v_mov_b32_e32 v80, v76 // 000000005AC0: 7EA0034C
|
|
v_cvt_f16_f32_e32 v80, v80 // 000000005AC4: 7EA01550
|
|
buffer_store_b16 v80, v89, s[16:19], 0 offen // 000000005AC8: E0640000 80445059
|
|
v_mul_f32_e32 v81, v92, v81 // 000000005AD0: 10A2A35C
|
|
v_add_f32_e32 v76, v91, v81 // 000000005AD4: 0698A35B
|
|
v_mov_b32_e32 v81, v76 // 000000005AD8: 7EA2034C
|
|
v_cvt_f16_f32_e32 v81, v81 // 000000005ADC: 7EA21551
|
|
buffer_store_b16 v81, v93, s[16:19], 0 offen // 000000005AE0: E0640000 8044515D
|
|
v_mul_f32_e32 v82, v84, v82 // 000000005AE8: 10A4A554
|
|
v_add_f32_e32 v76, v83, v82 // 000000005AEC: 0698A553
|
|
v_mov_b32_e32 v82, v76 // 000000005AF0: 7EA4034C
|
|
v_cvt_f16_f32_e32 v82, v82 // 000000005AF4: 7EA41552
|
|
buffer_store_b16 v82, v95, s[16:19], 0 offen // 000000005AF8: E0640000 8044525F
|
|
|
|
exit: // 0000000000005b00
|
|
s_endpgm // 000000005B00: BFB00000
|